# MLX: full text > MLX Network is a community building the largest Apple Silicon compute network, live on mainnet since 2026-10-18. MLX Node, a Mac menu bar app, runs AI work on Apple Silicon Macs and mines MLX tokens, and the owner controls when and how much each Mac works (MLX, the companion app for iPhone and iPad, App Store name MLX.app, controls them remotely). MLX Harness, a Mac app for running AI models locally, carries MLX Node inside. A Mac running MLX Node is a node of the network; without a dedicated seat it runs on a free seat. A dedicated seat (formerly called a node), matched to the Mac's chip, works like a stake: it reserves the Mac's place on the network and sets how fast it mines (it pays nothing up front since 2026-10-20); sending paid AI jobs to Macs with dedicated seats first is planned. Every block, transaction and seat is public and verifiable at explorer.mlx.app. MLX is not affiliated with Apple Inc. Key facts (from the live protocol code): - Total supply: 10,000,000,000 MLX, fixed at genesis. - Unit: $100. The network price follows a halving curve: what a unit buys at the curve price halves every 200,000 units, so the price doubles. Spot buys and sells are filled on this curve. Dedicated seats move the curve but pay nothing up front: they mine. - Dedicated seat tiers: two starter seats any Apple silicon Mac runs in full, then one per Mac chip class (same $100 per unit throughout; Latest editions are for Apple's newest chips, M6, M5 Pro, M5 Max, M5 Ultra: 1.5× the price and the weight): - Nano (starter seat: any Mac with Apple silicon, at full weight): $50, 0.5 units, mines up to 4 MLX a day, mined MLX locked 20 days. No Latest edition. - Lite (starter seat: any Mac with Apple silicon, at full weight): $100, 1 unit, mines up to 8 MLX a day, mined MLX locked 14 days. No Latest edition. - Core (Macs with an M-series chip): $250, 2.5 units, mines up to 20 MLX a day, mined MLX locked 10 days. Core Latest: $375, 3.75 units, up to 30 MLX a day. - Pro (Macs with a Pro or Max chip): $1,000, 10 units, mines up to 80 MLX a day, mined MLX locked 7 days. Pro Latest: $1,500, 15 units, up to 120 MLX a day. - Ultra (Macs with an Ultra chip): $4,000, 40 units, mines up to 320 MLX a day, mined MLX locked 3 days. Ultra Latest: $6,000, 60 units, up to 480 MLX a day. - Allocation: - Seat reserve: 30% (rewards.mlx) - Compute rewards: 35% (compute.mlx) - Community and referrals: 10% (community.mlx) - Treasury: 15% (treasury.mlx) - Contributors: 10% (team.mlx) - Compute rewards: since 2026-10-17, Macs running MLX Node earn MLX hourly from the compute pool, by availability, verified work and agreement with other Macs, weighted by seat units: each dedicated seat runs on one linked Mac and earns at the smaller of the seat's class and the Mac's (starter seats earn in full on any Mac); a wallet without a dedicated seat runs on a free seat at 0.125 units and is throttled: short AI jobs on a random heartbeat (a 1–9 second wait between jobs); dedicated seats get full jobs back to back, a steady flow. Capped at 8 MLX per unit per day. Mined MLX unlocks 30 days after the day it is earned on a free seat, or after the lock-in of the wallet's shortest-locked dedicated seat (3 to 20 days); USDC paid for jobs is never locked. - Mainnet: live since 2026-10-18. Mined MLX locks count from the day it is earned. Still to come: audits and slashing, paid jobs from clients (paid in USDC, never locked, sent to Macs with dedicated seats first), independent routers. - MLX for iPhone and iPad (App Store name MLX.app): the free companion to MLX Node. See the Macs linked to your wallet and what they mined, pause or resume them, set their limits, and get a notification when a Mac stops. It has no wallet features yet; sending and receiving come in a later version. - The network dollar is USDC; withdrawals are paid in USDC on Solana. - Contact: info@mlx.app (general), support@mlx.app (wallet, seats, payments, MLX Node, MLX Harness and MLX.app), contact@mlx.app (business, legal, security). Operator: MLX Foundation, Zug, Switzerland. - Languages: English, Chinese (Simplified), Chinese (Traditional), Hindi, Spanish, Arabic, Bengali, Portuguese, Russian, Japanese, German, French, Urdu, Indonesian, Turkish, Vietnamese, Korean, Italian, Persian, Thai, Polish, Ukrainian, Dutch, Malay, Filipino, Swahili, Tamil, Telugu, Marathi, Gujarati. This file is the full text of mlx.app for AI assistants and answer engines, built from the same sources as the pages on every deploy: the protocol constants, the compare catalogs and the English guides. The short index is https://mlx.app/llms.txt. Every outside number links its source; quote the page URL given with each section. ## Seats, mining and locks Source: the protocol code (lib/chain/curve.ts), the same constants as [Dedicated seats](https://mlx.app/nodes), [Rewards](https://mlx.app/rewards) and the [whitepaper](https://mlx.app/whitepaper). Caps are the most a seat can mine; what a Mac earns in an hour depends on its availability, verified work and agreement with other Macs. | Seat | Price | Units | Runs on | Mines up to (MLX a day) | MLX an hour | Mined MLX unlocks after (days) | | --- | --- | --- | --- | --- | --- | --- | | Nano | $50 | 0.5 | any Apple silicon Mac, in full | 4 | ~0.17 | 20 | | Lite | $100 | 1 | any Apple silicon Mac, in full | 8 | ~0.33 | 14 | | Core | $250 | 2.5 | Macs with an M-series chip | 20 | ~0.83 | 10 | | Core Latest | $375 | 3.75 | Macs with Apple's newest M chip | 30 | ~1.25 | 10 | | Pro | $1,000 | 10 | Macs with a Pro or Max chip | 80 | ~3.33 | 7 | | Pro Latest | $1,500 | 15 | Macs with an M5 Pro or M5 Max | 120 | ~5 | 7 | | Ultra | $4,000 | 40 | Macs with an Ultra chip | 320 | ~13.33 | 3 | | Ultra Latest | $6,000 | 60 | Macs with an M5 Ultra | 480 | ~20 | 3 | | Free seat (no dedicated seat) | $0 | 0.125 | any Apple silicon Mac running MLX Node, throttled | 1 | ~0.0417 | 30 | - κ = 8 MLX per seat unit per day; a unit is $100. Latest editions (for M6, M5 Pro, M5 Max, M5 Ultra) are 1.5× the price and the weight. - A dedicated seat runs on one linked Mac and earns at the smaller of the seat's class and the Mac's; starter seats (Nano, Lite) earn in full on any Mac. - A free seat is throttled: short AI jobs with a random 1–9 second wait between them; dedicated seats get full jobs back to back. - Mined MLX unlocks 30 days after the day it is earned on a free seat, or after the lock-in of the wallet's shortest-locked dedicated seat. USDC paid for AI jobs is never locked. - Since 2026-10-20 a dedicated seat pays nothing up front: it mines. ## Can a Mac run it? Local AI on Apple Silicon Every figure below is the compare engine's arithmetic on its sourced catalogs (models from the official model cards, chips from Apple's Tech Specs), the same as [mlx.app/compare](https://mlx.app/compare), [every Mac](https://mlx.app/mac) and [every model](https://mlx.app/run). Speeds are estimates, not measurements. Data checked 2026-10-19. Method: Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s [public Apple Silicon benchmarks](https://github.com/ggml-org/llama.cpp/discussions/4167) real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on [MLX Bench](https://mlx.app/bench). Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as [mlx.app/compare](https://mlx.app/compare). ### Best LLMs by memory size | Memory | Open models that run well | Biggest that runs well | Page | | --- | --- | --- | --- | | 8 GB | 18 of 81 | Granite 4.2 8B | https://mlx.app/mac/8gb | | 16 GB | 24 of 81 | Phi-4 Reasoning Vision 15B | https://mlx.app/mac/16gb | | 18 GB | 22 of 81 | Gemma 4 12B Unified | https://mlx.app/mac/18gb | | 24 GB | 31 of 81 | Raon-1 mini Preview | https://mlx.app/mac/24gb | | 32 GB | 43 of 81 | Qwen3.6 35B A3B | https://mlx.app/mac/32gb | | 36 GB | 44 of 81 | Recursive Flint | https://mlx.app/mac/36gb | | 48 GB | 44 of 81 | Recursive Flint | https://mlx.app/mac/48gb | | 64 GB | 45 of 81 | Nemotron 3.5 Bolt 74B A9B | https://mlx.app/mac/64gb | | 96 GB | 50 of 81 | gpt-oss-120b | https://mlx.app/mac/96gb | | 128 GB | 54 of 81 | Qwen3.5 122B A10B | https://mlx.app/mac/128gb | | 192 GB | 56 of 81 | MiniMax M2.7 | https://mlx.app/mac/192gb | | 256 GB | 61 of 81 | MiMo V2.6 Flash | https://mlx.app/mac/256gb | | 512 GB | 69 of 81 | Ling 3.1 Flash | https://mlx.app/mac/512gb | ### Every Mac, ranked for local AI Ranked by how many of 81 open models each runs well, then memory bandwidth. Speed column: gpt-oss-20b, best version that fits. | Mac | Run well | Bandwidth | Biggest that runs well | gpt-oss-20b | Page | | --- | --- | --- | --- | --- | --- | | M5 Ultra 512 GB | 69 | 1200 GB/s | Ling 3.1 Flash | 284–505 tok/s | https://mlx.app/mac/m5-ultra-512gb | | M3 Ultra 512 GB | 68 | 819 GB/s | Ling 3.1 Flash | 194–345 tok/s | https://mlx.app/mac/m3-ultra-512gb | | M5 Ultra 256 GB | 61 | 1200 GB/s | MiMo V2.6 Flash | 284–505 tok/s | https://mlx.app/mac/m5-ultra-256gb | | M3 Ultra 256 GB | 60 | 819 GB/s | MiMo V2.6 Flash | 194–345 tok/s | https://mlx.app/mac/m3-ultra-256gb | | M2 Ultra 192 GB | 56 | 800 GB/s | MiniMax M2.7 | 189–337 tok/s | https://mlx.app/mac/m2-ultra-192gb | | M2 Ultra 128 GB | 54 | 800 GB/s | Qwen3.5 122B A10B | 189–337 tok/s | https://mlx.app/mac/m2-ultra-128gb | | M1 Ultra 128 GB | 54 | 800 GB/s | Qwen3.5 122B A10B | 189–337 tok/s | https://mlx.app/mac/m1-ultra-128gb | | M5 Max 128 GB | 54 | 614 GB/s | Qwen3.5 122B A10B | 145–259 tok/s | https://mlx.app/mac/m5-max-128gb | | M4 Max 128 GB | 54 | 546 GB/s | Qwen3.5 122B A10B | 129–230 tok/s | https://mlx.app/mac/m4-max-128gb | | M3 Max 128 GB | 54 | 400 GB/s | Qwen3.5 122B A10B | 95–168 tok/s | https://mlx.app/mac/m3-max-128gb | | M5 Ultra 96 GB | 50 | 1200 GB/s | gpt-oss-120b | 284–505 tok/s | https://mlx.app/mac/m5-ultra-96gb | | M3 Ultra 96 GB | 49 | 819 GB/s | gpt-oss-120b | 194–345 tok/s | https://mlx.app/mac/m3-ultra-96gb | | M2 Max 96 GB | 49 | 400 GB/s | gpt-oss-120b | 95–168 tok/s | https://mlx.app/mac/m2-max-96gb | | M2 Ultra 64 GB | 45 | 800 GB/s | Nemotron 3.5 Bolt 74B A9B | 189–337 tok/s | https://mlx.app/mac/m2-ultra-64gb | | M1 Ultra 64 GB | 45 | 800 GB/s | Nemotron 3.5 Bolt 74B A9B | 189–337 tok/s | https://mlx.app/mac/m1-ultra-64gb | | M5 Max 64 GB | 45 | 614 GB/s | Nemotron 3.5 Bolt 74B A9B | 145–259 tok/s | https://mlx.app/mac/m5-max-64gb | | M4 Max 64 GB | 45 | 546 GB/s | Nemotron 3.5 Bolt 74B A9B | 129–230 tok/s | https://mlx.app/mac/m4-max-64gb | | M3 Max 64 GB | 45 | 400 GB/s | Nemotron 3.5 Bolt 74B A9B | 95–168 tok/s | https://mlx.app/mac/m3-max-64gb | | M2 Max 64 GB | 45 | 400 GB/s | Nemotron 3.5 Bolt 74B A9B | 95–168 tok/s | https://mlx.app/mac/m2-max-64gb | | M1 Max 64 GB | 45 | 400 GB/s | Nemotron 3.5 Bolt 74B A9B | 95–168 tok/s | https://mlx.app/mac/m1-max-64gb | | M5 Max 48 GB | 44 | 614 GB/s | Recursive Flint | 145–259 tok/s | https://mlx.app/mac/m5-max-48gb | | M4 Max 48 GB | 44 | 546 GB/s | Recursive Flint | 129–230 tok/s | https://mlx.app/mac/m4-max-48gb | | M5 Max 36 GB | 44 | 460 GB/s | Recursive Flint | 109–194 tok/s | https://mlx.app/mac/m5-max-36gb | | M4 Max 36 GB | 44 | 410 GB/s | Recursive Flint | 97–173 tok/s | https://mlx.app/mac/m4-max-36gb | | M3 Max 48 GB | 44 | 400 GB/s | Recursive Flint | 95–168 tok/s | https://mlx.app/mac/m3-max-48gb | | M3 Max 96 GB | 44 | 300 GB/s | gpt-oss-120b | 71–126 tok/s | https://mlx.app/mac/m3-max-96gb | | M2 Max 32 GB | 43 | 400 GB/s | Qwen3.6 35B A3B | 95–168 tok/s | https://mlx.app/mac/m2-max-32gb | | M1 Max 32 GB | 43 | 400 GB/s | Qwen3.6 35B A3B | 95–168 tok/s | https://mlx.app/mac/m1-max-32gb | | M5 Pro 64 GB | 40 | 307 GB/s | Nemotron 3.5 Bolt 74B A9B | 73–129 tok/s | https://mlx.app/mac/m5-pro-64gb | | M5 Pro 48 GB | 39 | 307 GB/s | Recursive Flint | 73–129 tok/s | https://mlx.app/mac/m5-pro-48gb | | M3 Max 36 GB | 39 | 300 GB/s | Recursive Flint | 71–126 tok/s | https://mlx.app/mac/m3-max-36gb | | M4 Pro 64 GB | 37 | 273 GB/s | Nemotron 3.5 Bolt 74B A9B | 65–115 tok/s | https://mlx.app/mac/m4-pro-64gb | | M4 Pro 48 GB | 36 | 273 GB/s | Recursive Flint | 65–115 tok/s | https://mlx.app/mac/m4-pro-48gb | | M2 Pro 32 GB | 35 | 200 GB/s | Qwen3.6 35B A3B | 47–84 tok/s | https://mlx.app/mac/m2-pro-32gb | | M1 Pro 32 GB | 35 | 200 GB/s | Qwen3.6 35B A3B | 47–84 tok/s | https://mlx.app/mac/m1-pro-32gb | | M6 32 GB | 34 | 170 GB/s | Qwen3.6 35B A3B | 40–72 tok/s | https://mlx.app/mac/m6-32gb | | M3 Pro 36 GB | 34 | 150 GB/s | Recursive Flint | 36–63 tok/s | https://mlx.app/mac/m3-pro-36gb | | M5 32 GB | 33 | 153 GB/s | Qwen3.6 35B A3B | 36–64 tok/s | https://mlx.app/mac/m5-32gb | | M4 32 GB | 32 | 120 GB/s | Qwen3.6 35B A3B | 28–51 tok/s | https://mlx.app/mac/m4-32gb | | M5 Pro 24 GB | 31 | 307 GB/s | Raon-1 mini Preview | 73–129 tok/s | https://mlx.app/mac/m5-pro-24gb | | M4 Pro 24 GB | 28 | 273 GB/s | Raon-1 mini Preview | 65–115 tok/s | https://mlx.app/mac/m4-pro-24gb | | M6 24 GB | 27 | 170 GB/s | Raon-1 mini Preview | 40–72 tok/s | https://mlx.app/mac/m6-24gb | | M5 24 GB | 26 | 153 GB/s | Raon-1 mini Preview | 36–64 tok/s | https://mlx.app/mac/m5-24gb | | M4 24 GB | 25 | 120 GB/s | Raon-1 mini Preview | 28–51 tok/s | https://mlx.app/mac/m4-24gb | | M2 Pro 16 GB | 24 | 200 GB/s | Phi-4 Reasoning Vision 15B | 56–100 tok/s | https://mlx.app/mac/m2-pro-16gb | | M1 Pro 16 GB | 24 | 200 GB/s | Phi-4 Reasoning Vision 15B | 56–100 tok/s | https://mlx.app/mac/m1-pro-16gb | | M3 24 GB | 23 | 100 GB/s | Raon-1 mini Preview | 24–42 tok/s | https://mlx.app/mac/m3-24gb | | M2 24 GB | 23 | 100 GB/s | Raon-1 mini Preview | 24–42 tok/s | https://mlx.app/mac/m2-24gb | | M6 16 GB | 22 | 153 GB/s | Gemma 4 12B Unified | 43–77 tok/s | https://mlx.app/mac/m6-16gb | | M5 16 GB | 22 | 153 GB/s | Gemma 4 12B Unified | 43–77 tok/s | https://mlx.app/mac/m5-16gb | | M3 Pro 18 GB | 22 | 150 GB/s | Gemma 4 12B Unified | 36–63 tok/s | https://mlx.app/mac/m3-pro-18gb | | M4 16 GB | 21 | 120 GB/s | Qwen3.5 9B | 34–60 tok/s | https://mlx.app/mac/m4-16gb | | M3 16 GB | 19 | 100 GB/s | Qwen3 8B | 28–50 tok/s | https://mlx.app/mac/m3-16gb | | M2 16 GB | 19 | 100 GB/s | Qwen3 8B | 28–50 tok/s | https://mlx.app/mac/m2-16gb | | M3 8 GB | 18 | 100 GB/s | Granite 4.2 8B | too big | https://mlx.app/mac/m3-8gb | | M2 8 GB | 18 | 100 GB/s | Granite 4.2 8B | too big | https://mlx.app/mac/m2-8gb | | M1 16 GB | 15 | 68 GB/s | Gemma 4 E4B | 19–34 tok/s | https://mlx.app/mac/m1-16gb | | M1 8 GB | 15 | 68 GB/s | Gemma 4 E4B | too big | https://mlx.app/mac/m1-8gb | ### Every model - [Can a Mac run DeepSeek V4.1 Flash 1017?](https://mlx.app/run/deepseek-v4.1-flash-1017): 4-bit needs about 335 GB; smallest Mac 512 GB; up to 60–107 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run DeepSeek V4 Pro 0813?](https://mlx.app/run/deepseek-v4-pro-0813): 4-bit needs about 972 GB; no single Mac holds it - [Can a Mac run DeepSeek V4 Flash 0731?](https://mlx.app/run/deepseek-v4-flash-0731): 4-bit needs about 173 GB; smallest Mac 256 GB; up to 74–132 tokens/s (estimate) on an M5 Ultra 256 GB - [Can a Mac run Qwen3.8 27B?](https://mlx.app/run/qwen3.8-27b): 4-bit needs about 16.7 GB; smallest Mac 24 GB; up to 19–33 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.8 Flash Next?](https://mlx.app/run/qwen3.8-flash-next): 4-bit needs about 109 GB; smallest Mac 192 GB; up to 110–196 tokens/s (estimate) on an M5 Ultra 256 GB - [Can a Mac run Qwen3.8 2.4T A95B?](https://mlx.app/run/qwen3.8-2.4t-a95b): 4-bit needs about 1.5 TB; no single Mac holds it - [Can a Mac run Qwen3.5 397B A17B?](https://mlx.app/run/qwen3.5-397b-a17b): 4-bit needs about 241 GB; smallest Mac 512 GB; up to 39–70 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Qwen3.5 122B A10B?](https://mlx.app/run/qwen3.5-122b-a10b): 4-bit needs about 74.1 GB; smallest Mac 128 GB; up to 64–114 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Ornith 1.5 122B?](https://mlx.app/run/ornith-1.5-122b): 4-bit needs about 74.1 GB; smallest Mac 128 GB; up to 64–114 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Qwen3.6 35B A3B?](https://mlx.app/run/qwen3.6-35b-a3b): 4-bit needs about 21.3 GB; smallest Mac 32 GB; up to 169–300 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.6 27B?](https://mlx.app/run/qwen3.6-27b): 4-bit needs about 16.7 GB; smallest Mac 24 GB; up to 19–33 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.5 9B?](https://mlx.app/run/qwen3.5-9b): 4-bit needs about 6.6 GB; smallest Mac 16 GB; up to 56–100 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.5 4B?](https://mlx.app/run/qwen3.5-4b): 4-bit needs about 3.8 GB; smallest Mac 8 GB; up to 126–223 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.5 2B?](https://mlx.app/run/qwen3.5-2b): 4-bit needs about 2.6 GB; smallest Mac 8 GB; up to 257–457 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3.5 0.8B?](https://mlx.app/run/qwen3.5-0.8b): 4-bit needs about 2 GB; smallest Mac 8 GB; up to 600–1067 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3 Coder Next?](https://mlx.app/run/qwen3-coder-next): 4-bit needs about 48.6 GB; smallest Mac 96 GB; up to 225–400 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3 Coder 30B A3B Instruct?](https://mlx.app/run/qwen3-coder-30b-a3b): 4-bit needs about 18.7 GB; smallest Mac 32 GB; up to 154–274 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3 8B?](https://mlx.app/run/qwen3-8b): 4-bit needs about 6.1 GB; smallest Mac 16 GB; up to 62–110 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Qwen3 0.6B?](https://mlx.app/run/qwen3-0.6b): 4-bit needs about 1.8 GB; smallest Mac 8 GB; up to 900–1600 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run gpt-oss-20b?](https://mlx.app/run/gpt-oss-20b): native 4.25-bit needs about 15.3 GB; smallest Mac 24 GB; up to 284–505 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run gpt-oss-120b?](https://mlx.app/run/gpt-oss-120b): native 4.25-bit needs about 70.5 GB; smallest Mac 96 GB; up to 200–356 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Llama 3.1 8B Instruct?](https://mlx.app/run/llama-3.1-8b): 4-bit needs about 6 GB; smallest Mac 8 GB; up to 64–113 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Llama 3.3 70B Instruct?](https://mlx.app/run/llama-3.3-70b): 4-bit needs about 42.6 GB; smallest Mac 64 GB; up to 14–24 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Llama 4 Scout 17B-16E Instruct?](https://mlx.app/run/llama-4-scout): 4-bit needs about 66.2 GB; smallest Mac 96 GB; up to 56–100 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Llama 4 Maverick 17B-128E Instruct?](https://mlx.app/run/llama-4-maverick): 4-bit needs about 243 GB; smallest Mac 512 GB; up to 39–70 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Llama 3.2 1B Instruct?](https://mlx.app/run/llama-3.2-1b): 4-bit needs about 2.2 GB; smallest Mac 8 GB; up to 415–738 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Llama 3.2 3B Instruct?](https://mlx.app/run/llama-3.2-3b): 4-bit needs about 3.3 GB; smallest Mac 8 GB; up to 159–282 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Muse Spark Open?](https://mlx.app/run/muse-spark-open): 4-bit needs about 496 GB; smallest Mac 512 GB; up to 41–73 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Muse Glimmer 30B?](https://mlx.app/run/muse-glimmer-30b): 4-bit needs about 18.2 GB; smallest Mac 32 GB; up to 17–30 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Gemma 4 31B?](https://mlx.app/run/gemma-4-31b): 4-bit needs about 18.8 GB; smallest Mac 32 GB; up to 17–29 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Gemma 4 26B A4B?](https://mlx.app/run/gemma-4-26b-a4b): 4-bit needs about 15.7 GB; smallest Mac 24 GB; up to 135–240 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Gemma 4 12B Unified?](https://mlx.app/run/gemma-4-12b): 4-bit needs about 8.2 GB; smallest Mac 16 GB; up to 43–76 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Gemma 4 E4B?](https://mlx.app/run/gemma-4-e4b): 4-bit needs about 6 GB; smallest Mac 8 GB; up to 113–200 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Gemma 4 E2B?](https://mlx.app/run/gemma-4-e2b): 4-bit needs about 4.4 GB; smallest Mac 8 GB; up to 225–400 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Mistral Small 4 (119B)?](https://mlx.app/run/mistral-small-4): 4-bit needs about 72.3 GB; smallest Mac 128 GB; up to 97–173 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Mistral Medium 3.5 (128B)?](https://mlx.app/run/mistral-medium-3.5): 4-bit needs about 77.8 GB; smallest Mac 128 GB; up to 5–9 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Mistral Large 3 (675B)?](https://mlx.app/run/mistral-large-3): 4-bit needs about 410 GB; smallest Mac 512 GB; up to 30–54 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Devstral Small 2 (24B)?](https://mlx.app/run/devstral-small-2): 4-bit needs about 15 GB; smallest Mac 24 GB; up to 21–38 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Devstral 2 (123B)?](https://mlx.app/run/devstral-2-123b): 4-bit needs about 74.7 GB; smallest Mac 128 GB; up to 5–9 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Magistral Small 1.2 (2509)?](https://mlx.app/run/magistral-small-2509): 4-bit needs about 15 GB; smallest Mac 24 GB; up to 21–38 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Ministral 3 8B?](https://mlx.app/run/ministral-3-8b): 4-bit needs about 6.5 GB; smallest Mac 16 GB; up to 57–101 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Mistral Small 3.2 (24B)?](https://mlx.app/run/mistral-small-3.2): 4-bit needs about 15 GB; smallest Mac 24 GB; up to 21–38 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Phi-4 (14B)?](https://mlx.app/run/phi-4): 4-bit needs about 9.4 GB; smallest Mac 16 GB; up to 36–64 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Phi-4 mini instruct?](https://mlx.app/run/phi-4-mini): 4-bit needs about 3.6 GB; smallest Mac 8 GB; up to 135–240 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Phi-4 Reasoning Vision 15B?](https://mlx.app/run/phi-4-reasoning-vision-15b): 4-bit needs about 9.9 GB; smallest Mac 16 GB; up to 34–60 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Kimi K3?](https://mlx.app/run/kimi-k3): native 4.25-bit needs about 1.7 TB; no single Mac holds it - [Can a Mac run Kimi K2.6?](https://mlx.app/run/kimi-k2.6): native 4-bit needs about 643 GB; smallest Mac 512 GB; up to 54–96 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Kimi K2.7 Code?](https://mlx.app/run/kimi-k2.7-code): native 4-bit needs about 643 GB; smallest Mac 512 GB; up to 54–96 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run GLM-5.3?](https://mlx.app/run/glm-5.3): 4-bit needs about 452 GB; smallest Mac 512 GB; up to 31–55 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run GLM-5.3 Flash?](https://mlx.app/run/glm-5.3-flash): 4-bit needs about 194 GB; smallest Mac 512 GB; up to 28–50 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run GLM-4.7 Flash?](https://mlx.app/run/glm-4.7-flash): 4-bit needs about 18.4 GB; smallest Mac 32 GB; up to 169–300 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run GLM-4.5 Air?](https://mlx.app/run/glm-4.5-air): 4-bit needs about 64.4 GB; smallest Mac 96 GB; up to 79–141 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run MiniMax M3?](https://mlx.app/run/minimax-m3): 4-bit needs about 260 GB; smallest Mac 512 GB; up to 29–51 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run MiniMax M2.7?](https://mlx.app/run/minimax-m2.7): 4-bit needs about 140 GB; smallest Mac 192 GB; up to 96–171 tokens/s (estimate) on an M5 Ultra 256 GB - [Can a Mac run Nemotron 3 Nano 30B A3B?](https://mlx.app/run/nemotron-3-nano-30b-a3b): 4-bit needs about 18.4 GB; smallest Mac 32 GB; up to 146–259 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Nemotron 3.5 Lightning 30B A3B?](https://mlx.app/run/nemotron-3.5-lightning-30b-a3b): 4-bit needs about 18.4 GB; smallest Mac 32 GB; up to 169–300 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Nemotron 3 Super 120B A12B?](https://mlx.app/run/nemotron-3-super-120b-a12b): 4-bit needs about 72.9 GB; smallest Mac 128 GB; up to 53–94 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Nemotron 3.5 Bolt 74B A9B?](https://mlx.app/run/nemotron-3.5-bolt-74b-a9b): 4-bit needs about 44.9 GB; smallest Mac 64 GB; up to 74–132 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Nemotron 3 Ultra 550B A55B?](https://mlx.app/run/nemotron-3-ultra-550b-a55b): 4-bit needs about 334 GB; smallest Mac 512 GB; up to 17–31 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Nemotron 3 Nano 4B?](https://mlx.app/run/nemotron-3-nano-4b): 4-bit needs about 3.7 GB; smallest Mac 8 GB; up to 129–229 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Granite 4.2 8B?](https://mlx.app/run/granite-4.2-8b): 4-bit needs about 6 GB; smallest Mac 8 GB; up to 64–113 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Granite 4.2 3B?](https://mlx.app/run/granite-4.2-3b): 4-bit needs about 3.2 GB; smallest Mac 8 GB; up to 169–300 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Granite 4.2 30B?](https://mlx.app/run/granite-4.2-30b): 4-bit needs about 18.4 GB; smallest Mac 32 GB; up to 17–30 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run SmolLM3 3B?](https://mlx.app/run/smollm3-3b): 4-bit needs about 3.2 GB; smallest Mac 8 GB; up to 169–300 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run LFM2.5 24B A2B?](https://mlx.app/run/lfm2.5-24b-a2b): 4-bit needs about 15 GB; smallest Mac 24 GB; up to 257–457 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run LFM2.5 2.6B?](https://mlx.app/run/lfm2.5-2.6b): 4-bit needs about 3 GB; smallest Mac 8 GB; up to 186–331 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run MiniCPM5 2B?](https://mlx.app/run/minicpm5-2b): 4-bit needs about 2.9 GB; smallest Mac 8 GB; up to 200–356 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Hy3?](https://mlx.app/run/hy3): 4-bit needs about 179 GB; smallest Mac 256 GB; up to 46–81 tokens/s (estimate) on an M5 Ultra 256 GB - [Can a Mac run MiMo V2.6 Flash?](https://mlx.app/run/mimo-v2.6-flash): native 4.48-bit needs about 187 GB; smallest Mac 256 GB; up to 64–114 tokens/s (estimate) on an M5 Ultra 256 GB - [Can a Mac run MiMo V2.6 Pro?](https://mlx.app/run/mimo-v2.6-pro): native 4.44-bit needs about 611 GB; smallest Mac 512 GB; up to 41–73 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Raon-1 mini Preview?](https://mlx.app/run/raon-1-mini-preview): 4-bit needs about 17.7 GB; smallest Mac 24 GB; up to 117–209 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Recursive Flint?](https://mlx.app/run/flint): 4-bit needs about 24.2 GB; smallest Mac 36 GB; up to 200–356 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Hy-Memory-30B-A3B?](https://mlx.app/run/hy-memory-30b-a3b): 4-bit needs about 18.4 GB; smallest Mac 32 GB; up to 159–282 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Ling 3.1 Flash?](https://mlx.app/run/ling-3.1-flash): 4-bit needs about 340 GB; smallest Mac 512 GB; up to 38–68 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Nemotron-Labs Dory 1.0 4B?](https://mlx.app/run/nemotron-labs-dory-1.0-4b): 4-bit needs about 3.7 GB; smallest Mac 8 GB; up to 87–155 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Nemotron 3.5 Super VL 120B A12B?](https://mlx.app/run/nemotron-3.5-super-vl-120b-a12b): 4-bit needs about 72.9 GB; smallest Mac 128 GB; up to 53–94 tokens/s (estimate) on an M2 Ultra 128 GB - [Can a Mac run Sage-A1.5?](https://mlx.app/run/sage-a1.5): 4-bit needs about 21.3 GB; smallest Mac 32 GB; up to 174–310 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Spark-X2.6 4B?](https://mlx.app/run/spark-x2.6-4b): 4-bit needs about 3.8 GB; smallest Mac 8 GB; up to 123–218 tokens/s (estimate) on an M5 Ultra 96 GB - [Can a Mac run Beam 1.0?](https://mlx.app/run/beam-1.0): 4-bit needs about 304 GB; smallest Mac 512 GB; up to 42–74 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Inkling?](https://mlx.app/run/inkling): 4-bit needs about 592 GB; smallest Mac 512 GB; up to 42–75 tokens/s (estimate) on an M5 Ultra 512 GB - [Can a Mac run Inkling Small?](https://mlx.app/run/inkling-small): 4-bit needs about 168 GB; smallest Mac 256 GB; up to 79–141 tokens/s (estimate) on an M5 Ultra 256 GB ## Open data [mlx.app/data](https://mlx.app/data): the data behind the tables above as free downloads, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), version 2026-10-19. File list with sizes and SHA-256: https://mlx.app/data/index.json. The fit table is estimated with the method above (https://mlx.app/compare#cp-how). - **Apple Silicon Mac configurations** (58 rows): Every Apple Silicon chip from M1 to M6 in every memory size Apple sells it with: unified memory, memory bandwidth, GPU core options and the Macs it ships in, read from Apple's Tech Specs pages, with the source of every row. https://mlx.app/data/macs.csv · https://mlx.app/data/macs.json - **Open AI models: parameters, active parameters and memory** (82 rows): Every open-weight language model on mlx.app/compare: total and active parameters, architecture (dense or mixture of experts), context length, licence, release date, precision and weight size, read from each model's official Hugging Face card, config.json and file listing. https://mlx.app/data/models.csv · https://mlx.app/data/models.json - **Which open model runs on which Mac, and how fast (estimates)** (4,698 rows): Every Apple Silicon Mac configuration against every open model: the highest-precision version that fits in the memory macOS gives the GPU, how much memory it needs with room for context, and the estimated writing speed range in tokens a second. Estimates from published specs, not measurements; the method is at mlx.app/compare. https://mlx.app/data/fit.csv · https://mlx.app/data/fit.json ## Every Apple Silicon chip Apple's figures for each chip's largest memory configuration (lower configurations can have less), from [mlx.app/learn/chip](https://mlx.app/learn/chip). Updated 2026-10-19. | Chip | Memory up to | Bandwidth | Macs | Page | | --- | --- | --- | --- | --- | | Apple M1 | 16 GB | 68 GB/s | MacBook Air (2020), 13-inch MacBook Pro (2020), Mac mini (2020), iMac (2021) | https://mlx.app/learn/chip/m1 | | Apple M1 Pro | 32 GB | 200 GB/s | 14- and 16-inch MacBook Pro (2021) | https://mlx.app/learn/chip/m1-pro | | Apple M1 Max | 64 GB | 400 GB/s | 14- and 16-inch MacBook Pro (2021), Mac Studio (2022) | https://mlx.app/learn/chip/m1-max | | Apple M1 Ultra | 128 GB | 800 GB/s | Mac Studio (2022) | https://mlx.app/learn/chip/m1-ultra | | Apple M2 | 24 GB | 100 GB/s | MacBook Air (2022, 2023), 13-inch MacBook Pro (2022), Mac mini (2023) | https://mlx.app/learn/chip/m2 | | Apple M2 Pro | 32 GB | 200 GB/s | 14- and 16-inch MacBook Pro (2023), Mac mini (2023) | https://mlx.app/learn/chip/m2-pro | | Apple M2 Max | 96 GB | 400 GB/s | 14- and 16-inch MacBook Pro (2023), Mac Studio (2023) | https://mlx.app/learn/chip/m2-max | | Apple M2 Ultra | 192 GB | 800 GB/s | Mac Studio (2023), Mac Pro (2023) | https://mlx.app/learn/chip/m2-ultra | | Apple M3 | 24 GB | 100 GB/s | MacBook Air (2024), 14-inch MacBook Pro (2023), iMac (2023) | https://mlx.app/learn/chip/m3 | | Apple M3 Pro | 36 GB | 150 GB/s | 14- and 16-inch MacBook Pro (2023) | https://mlx.app/learn/chip/m3-pro | | Apple M3 Max | 128 GB | 300–400 GB/s | 14- and 16-inch MacBook Pro (2023) | https://mlx.app/learn/chip/m3-max | | Apple M3 Ultra | 512 GB | 819 GB/s | Mac Studio (2025) | https://mlx.app/learn/chip/m3-ultra | | Apple M4 | 32 GB | 120 GB/s | MacBook Air (2025), 14-inch MacBook Pro (2024), Mac mini (2024), iMac (2024) | https://mlx.app/learn/chip/m4 | | Apple M4 Pro | 64 GB | 273 GB/s | Mac mini (2024), 14- and 16-inch MacBook Pro (2024) | https://mlx.app/learn/chip/m4-pro | | Apple M4 Max | 128 GB | 410–546 GB/s | 14- and 16-inch MacBook Pro (2024), Mac Studio (2025) | https://mlx.app/learn/chip/m4-max | | Apple M5 | 32 GB | 153 GB/s | 14-inch MacBook Pro (2025), MacBook Air (2026) | https://mlx.app/learn/chip/m5 | | Apple M5 Pro | 64 GB | 307 GB/s | 14- and 16-inch MacBook Pro (2026), Mac mini (2026) | https://mlx.app/learn/chip/m5-pro | | Apple M5 Max | 128 GB | 460–614 GB/s | 14- and 16-inch MacBook Pro (2026), Mac Studio (2026) | https://mlx.app/learn/chip/m5-max | | Apple M5 Ultra | 512 GB | 1200 GB/s | Mac Studio (2026) | https://mlx.app/learn/chip/m5-ultra | | Apple M6 | 32 GB | 153–170 GB/s | Mac mini (2026) | https://mlx.app/learn/chip/m6 | ## Guides Every guide at [mlx.app/learn](https://mlx.app/learn), in full (English; 30 languages on the site). ### Start earning MLX on your Mac URL: https://mlx.app/learn/start-earning-mlx-on-mac · Updated 2026-10-19 · Get started **The short version** - Install MLX Node, link it to your MLX wallet and set your limits. Setup takes a few minutes and costs nothing. - Your Mac runs small open language models with Apple's MLX. Other Macs check the answers, and rewards are worked out every hour. - On a free seat, a wallet mines up to 1 MLX a day. A dedicated seat raises that ceiling. These are limits, not promises. - Mined MLX is locked for 30 days from the day it's earned, or 3 to 20 days with a dedicated seat. Your wallet counts it up live while your Mac works. MLX Node lends your Mac to MLX Network: it runs small AI jobs, other Macs check the answers, and the network pays MLX Tokens every hour. Here is the whole setup, the settings that keep you in control, and how to see what your Mac has earned. #### What you need - A Mac with Apple Silicon (M1 or later) running macOS 14 or later. Intel Macs aren't supported. - An internet connection. By default MLX Node only works while the Mac is plugged in. - An MLX wallet. You can create one during setup, or first follow [Create an MLX wallet](https://mlx.app/learn/create-mlx-wallet). - Nothing to buy. The app is free, and a dedicated seat is optional. #### Install and set up MLX Node MLX Node lives in the menu bar. Everything happens in one panel that opens from its icon, and the first run walks you through five short screens. 1. [Download MLX Node](https://mlx.app/download), move it to your Applications folder and open it. Click its icon in the menu bar, then **Get started**. 2. **Connect your wallet.** Press **Connect** and your MLX wallet opens right in the panel, or choose **Use my browser instead**. Sign in, or create a wallet. Unless you've just created it, the wallet asks you to confirm it's you with your passkey or recovery phrase. Then approve the Mac with **Connect to** your wallet's name. 3. **You're in charge.** Set your limits (see below) and press **Continue**. You can change them any time in Settings. 4. **Measure this Mac.** **Run the test** downloads a small model once (about 300 MB) and measures how many tokens a second your Mac writes. You can **Skip for now**. 5. **All set.** Press **Done**. The menu bar icon turns orange while MLX Node is running. > Note: Linking needs two signatures: the Mac signs the request with its own key, kept in its Secure Enclave, and your wallet signs the approval. A Mac works for one wallet at a time; connecting it to another wallet moves it. #### Choose when your Mac works MLX Node works only when every one of your limits is met, and stops as soon as one isn't. From MLX Node 0.8.4. MLX Node also waits while Low Power Mode is on. | Setting | What it does | New installs | | --- | --- | --- | | **After idle for** | Off: you start and stop MLX Node yourself, and it works even while you use the Mac. 1 to 60 minutes: it starts after the Mac has been idle that long and stops the moment you're back. | Off | | **Only when plugged in** | Waits for power, and stops if you unplug a laptop. | On | | **Memory MLX may use** | The share of the Mac's memory a job may take, from 20% to 75%. | 50% | | **Hours** | **Any time**, or **Only at night** between the hours you pick. | Any time | | **Stop when the Mac is** | **Warm** or **Hot**: MLX Node lets the Mac cool down before it carries on. | Warm | With **After idle for** off, the panel shows **Start** and **Stop**: once your Mac is linked, MLX Node works within your other limits until you press **Stop**. With an idle time set, it shows **Pause** and **Resume** instead. To stop for good, choose **Quit MLX Node**. #### What your Mac runs Jobs are prompts for small open language models, most of them [Qwen2.5-0.5B-Instruct in 4-bit](https://huggingface.co/mlx-community/Qwen2.5-0.5B-Instruct-4bit), run with Apple's MLX framework. The answers go back to the network, which compares them with other Macs' answers to the same prompts and with known answers. The panel shows what's happening: the live speed in tokens a second, **Jobs today**, **Checks passed** (how often your Mac's answers matched other Macs'), what your wallet mined today and what's locked. The app is signed, notarized by Apple and sandboxed: it can't read your files, photos or messages. [Is MLX safe?](https://mlx.app/learn/is-mlx-safe) explains what it can see and what it sends. #### How rewards are worked out Every hour (UTC), each linked Mac gets a score from three things multiplied together: - **Availability:** the minutes it was ready to work or working, out of 60. - **Delivery:** verified jobs it finished, against one job every two of those minutes. - **Quality:** once it has been compared at least 10 times, the share of its answers that agreed with other Macs'. The network releases a small share of its compute pool each hour and divides it by score and weight, with a ceiling per wallet. A wallet without a dedicated seat earns from its best Mac only, on a free seat; free seats share at most 10% of each hour. Each dedicated seat runs on one of your MLX Nodes, the biggest seat on the biggest Mac, and earns at the smaller of the two classes. A seat works like a stake: its units set your share of each hour's rewards. Every linked Mac gets the same jobs today; paid AI jobs, when they arrive, are planned to go to Macs with dedicated seats first. Ceilings from the protocol; a seat running on a smaller Mac earns at that Mac's class. What you actually get depends on the hours your Mac works, its score and how many Macs share each hour. | Your wallet has | Runs at full rate on | Mines up to, per day | | --- | --- | --- | | A free seat only | Any Apple Silicon Mac | 1 MLX | | A Nano seat | Any Apple Silicon Mac | 4 MLX | | A Lite seat | Any Apple Silicon Mac | 8 MLX | | A Core seat | Any Apple Silicon Mac | 20 MLX | | A Pro seat | A Mac with a Pro or Max chip | 80 MLX | | An Ultra seat | A Mac with an Ultra chip | 320 MLX | Each hour's reward is paid in one public transaction from compute.mlx and added to that day's lock. Mined MLX unlocks 30 days after the day it was earned on a free seat, or after your seat's lock-in (3 to 20 days), and then you can send or sell it. MLX Tokens can fall in value or become worthless, so read the [risks](https://mlx.app/risks) before you buy a [dedicated seat](https://mlx.app/nodes). #### Watch what you've earned - **In the panel:** **Wallet mined today**, the locked total and the most your Mac can earn a day where it is. - **In your wallet:** while a Mac works, the **Locked MLX** row counts up live and reads **Mining** with the hourly rate; it says **Paused** when your Macs stop. Open it to see when each day's MLX unlocks. - **In Activity:** one **Mined on your Macs** row per day. - **On the explorer:** every hourly reward is a public transaction on [explorer.mlx.app](https://explorer.mlx.app), and your Mac appears on [MLX World](https://world.mlx.app) at its approximate city. - **On iPhone and iPad:** MLX.app lists your Macs and what they mined today, pauses or resumes them, edits their limits and tells you when one stops. Its App Store status is on the [download page](https://mlx.app/download). #### Questions **How much MLX can my Mac mine a day?** Up to 1 MLX a day for a wallet on a free seat, and up to 20 with a Core seat, more with Pro and Ultra seats running on Macs of their class. These are ceilings: the real amount depends on how long your Mac works and how well it scores. **Do I need a dedicated seat to mine MLX?** No. MLX Node is free and mines on a free seat. A dedicated seat is optional: like a stake, it raises how much a Mac of its class can earn. It pays nothing up front, and what your Mac mines with it is locked for 3 to 20 days: the smaller the seat, the longer the lock. **Do more Macs earn more?** On a free seat, no: a wallet earns from its best Mac only. Each dedicated seat runs on one linked Mac, so a wallet with several seats needs a Mac for each to earn in full. **When can I sell the MLX I mine?** Each day's mining unlocks 30 days after that day on a free seat, or after your seat's lock-in (3 to 20 days) with a dedicated seat. Locked MLX counts toward your total but can't be sent or sold before its date. **Will MLX Node slow my Mac down while I use it?** With After idle for set, it stops the moment you touch the Mac. In manual mode it keeps working until you press Stop, within the memory limit you chose. ### Create your MLX wallet URL: https://mlx.app/learn/create-mlx-wallet · Updated 2026-10-19 · Get started **The short version** - Open wallet.mlx.app, choose a name and press Create with passkey. Face ID or Touch ID saves the passkey on your device. - Your name, like yourname.mlx, is your public address. It's first come, first served. - Back up the 24-word recovery phrase straight away. If you lose both your passkey and the phrase, no one can recover the wallet. - Use the same wallet on your other devices through your synced passkey, or restore it with the phrase. An MLX wallet holds your MLX Tokens, USDC and dedicated seats, and it's what your Macs earn for. It takes a name and a passkey, no email and no seed phrase to start. Here's every step, and the one thing you must not skip. #### What you need - A browser that can create passkeys: Safari on iPhone, iPad or Mac, or Chrome. - Face ID, Touch ID or your device passcode, which unlock the passkey. - A passkey manager such as iCloud Keychain or Google Password Manager, which keeps the passkey and syncs it to your other devices. - Pen and paper for the recovery phrase. #### Create your wallet, step by step 1. Open [wallet.mlx.app](https://wallet.mlx.app) and choose **New wallet**. 2. Type a name under **Account name**. Use 2 to 32 lowercase letters, numbers and hyphens, starting and ending with a letter or number. The wallet adds .mlx and tells you when the name is available. A few names, like treasury or support, are reserved. 3. Press **Create with passkey** and confirm with Face ID or Touch ID. Your device saves a passkey for mlx.app. 4. The wallet opens and greets you by name. Creating it accepts the [Terms](https://mlx.app/terms) and [Privacy Policy](https://mlx.app/privacy). Names are public and can't be changed later, so pick one you're happy to share: people send you MLX and USDC by typing it. #### What the passkey does A passkey is a sign-in credential your device stores in its passkey manager. On Apple devices it lives in the Passwords app, syncs through iCloud Keychain, which is end-to-end encrypted, and is unlocked with Face ID or Touch ID ([Apple](https://support.apple.com/guide/platform-support/use-passkeys-in-iphone-and-ipad-sup742154734/web)). For MLX, the passkey carries your wallet's secret: 32 random bytes made on your device when you create the wallet. Each time you unlock, the browser hands that secret to the wallet page only after Face ID or Touch ID, and the wallet derives your signing key from it on your device. mlx.app's servers never receive it. [How the MLX wallet works](https://mlx.app/learn/how-the-mlx-wallet-works) has the details. #### Back up your recovery phrase The same secret can be written as 24 words. They restore your wallet on any device, even without your passkey, so they're your safety net if you lose your phone or your passkey manager account. 1. Open **Settings**: the cog at the top right, which carries an orange dot on a phone until you've backed up. Under **Security**, the **Recovery phrase** row says **Not backed up yet**. 2. Press **Back up**, make sure no one can see your screen, and press **Reveal phrase**. 3. Write the 24 words down in order and keep them offline. Then press **I've written it down**. 4. Type the words the wallet asks for from your copy and press **Confirm**. The row now says it's backed up, and you can view the phrase again any time. > Warning: Anyone with your 24 words controls your wallet. MLX will never ask for them, and can't recover a wallet if you lose both the passkey and the phrase. The wallet remembers the backup in the browser where you made it, so another browser may still say **Not backed up yet**. Checking your copy there clears it. #### Auto-lock and locking While the wallet is unlocked, its key is kept in that browser tab's session memory. Closing the tab clears it, and so does locking. - **Auto-lock** in Settings locks the wallet after a stretch without activity: 1 hour, 1 day, 30 days (the default) or Never. - **Lock wallet** removes the key from the tab now. Unlock again with **Continue with passkey**. - **Test my passkey** checks that Face ID or Touch ID on this device opens this wallet. #### Use it on another device, and link a Mac ##### On another phone, tablet or computer Open [wallet.mlx.app](https://wallet.mlx.app), choose **I have one** and **Continue with passkey**. If your passkey has synced to that device, you're in. If not, choose **Use recovery phrase**, type your 24 words and press **Restore wallet**; then you can add a passkey on that device from **Test my passkey** in Settings. MLX.app for iPhone and iPad signs in with the same passkey or recovery phrase, and keeps the key in that device's Keychain. ##### Link a Mac Install [MLX Node](https://mlx.app/download) and press **Connect** in its panel. Sign in to your wallet, confirm it's you, and approve the Mac with **Connect to** your name. Your linked Macs are listed under **Macs** in Settings, where you can unlink one. [Start earning MLX on your Mac](https://mlx.app/learn/start-earning-mlx-on-mac) walks through the rest. #### Questions **Does an MLX wallet cost anything?** No. Creating a wallet is free, and sending MLX or USDC to another MLX wallet is instant and free. Withdrawing USDC to Solana has a 1 USDC network fee. **Do I need an email or a seed phrase to start?** No. A name and a passkey are enough. The 24-word recovery phrase is created with the wallet; back it up from Settings right after. **What happens if I lose my phone?** If your passkey syncs through iCloud Keychain or Google Password Manager, sign in on another device. If not, restore with your recovery phrase. Without either, the wallet can't be recovered by anyone. **Can I change my wallet name?** No. A name is fixed once the account is created. If you want a different one, create a new wallet and send your balance across. **Can MLX reset my passkey or recover my wallet?** No. MLX never receives your passkey, wallet secret or recovery phrase, so it can't recover the wallet or move anything for you. ### How to use MLX Harness URL: https://mlx.app/learn/how-to-use-mlx-harness · Updated 2026-10-19 · Get started **The short version** - MLX Harness runs AI models on your Mac. Prompts and answers stay on the Mac unless you connect a remote endpoint yourself. - It's built on vMLX by JANGQ AI and licensed under Apache 2.0. Download it at [harness.mlx.app](https://harness.mlx.app) or [mlx.app/download](https://mlx.app/download). - Four pages: Chat & Images, Servers & API, Models and Network, with Preferences at the bottom of the sidebar. - Every running model is served on a local API at localhost:8080 in OpenAI, Anthropic and Ollama formats, so apps and coding agents can use it. MLX Harness is our free Mac app for running language, vision and image models on your own Mac with Apple's MLX framework. It's free to download. Here is what it does, page by page. #### Install MLX Harness > Note: Download MLX Harness at [harness.mlx.app](https://harness.mlx.app). It's signed and notarized by Apple, carries MLX Node inside, and tells you in the app when a new version is out. You'll need a Mac with Apple Silicon and macOS 14.5 or later. The inference engine ships inside the app, so there's nothing else to install: download it, move MLX Harness to Applications and open it. MLX Harness is a modified version of [vMLX](https://github.com/jjang-ai/vmlx) by JANGQ AI, released under the Apache License 2.0. The engine underneath is vMLX's, unchanged; MLX Harness changes the app around it and adds the Network page. #### Pick and download a model Models come from [Hugging Face](https://huggingface.co/models), most of them in the MLX format from the [mlx-community](https://huggingface.co/mlx-community) organization. Before you download, check what fits: [Compare](https://mlx.app/compare) shows which models your Mac's memory holds and roughly how fast each writes. 1. Open **Models** and choose **Find & download**. Search, and sort by downloads, trending, likes, recently updated or model size; the search lists MLX models. 2. Click a model to read its README, then download it. Downloads run in the background, and **Downloads** shows their progress. 3. For gated models such as Llama, add your Hugging Face token in the **HF Token:** field first, and accept the model's licence on huggingface.co. 4. **Library & tools** lists the models on your Mac. **Inspect Model** shows a model's metadata, **Diagnose Model** runs health checks and **Convert Model** turns a Hugging Face model into a quantized MLX one: **Balanced (4-bit)**, **Quality (8-bit)** or **Compact (3-bit)**. Downloads go to your Hugging Face cache folder unless you choose another under **Download to:**. MLX Harness also finds models already in that cache, in ~/.mlxstudio/models and in folders you add under **Directories**. #### Chat and images ##### Chat and reasoning 1. Open **Chat & Images**, pick a model in the chat toolbar's model menu, or **Add local model**. It starts with its detected defaults; advanced options stay in **Servers & API**. 2. Type a message. Reasoning models show their thinking in a separate box. **Enable Thinking** in the chat settings sets it to Auto, On or Off, and **Reasoning Effort** sets its depth for models that support it. 3. Vision models take images: drop a file on the chat or use the attach button. Video, audio and text files work with models that accept them. 4. Chats are saved in the sidebar and searchable. Settings are saved per model. You can also chat with a model somewhere else: **Connect remote endpoint** takes any OpenAI-compatible API address. Only then do your prompts leave the Mac. ##### Generate and edit images In **Chat & Images**, choose **Image**. Image models run through [mflux](https://github.com/filipstrand/mflux), and the image server loads a model from a folder on your Mac: **Find mflux models** opens the image search in **Models**, then **Load selected folder** starts it. Families MLX Harness can load. Memory varies by model and precision; the app suggests Flux Schnell or Klein 4B at 4-bit for a 16 GB Mac, and 8-bit or full precision for editing models. | Use | Model families | | --- | --- | | Generate | Flux Schnell, Flux Dev, Z-Image Turbo, FLUX.2 Klein 4B and 9B, Qwen Image, Qwen Image 2.1, FIBO, Ideogram 4 | | Edit | Qwen Image Edit, Qwen Image 2.1, Flux Kontext, FLUX.2 Klein Edit, Flux Fill (inpainting with a mask you paint) | | Upscale | SeedVR2 3B | Some image models, such as Flux Dev, are gated on Hugging Face and need your token and an accepted licence before they download. #### Agents and tools MLX Harness can let a model act on your Mac. In the chat settings, turn on **Enable Built-in Coding Tools** and choose a **Working Directory**: every file operation and command runs inside it. Then switch tool categories on or off: From MLX Harness's built-in tool list. MCP servers add their own tools next to these. | Category | Built-in tools | | --- | --- | | **File I/O** | `read_file`, `write_file`, `edit_file`, `patch_file`, `batch_edit`, `insert_text`, `replace_lines`, `apply_regex`, `copy_file`, `move_file`, `delete_file`, `create_directory`, `list_directory`, `read_image`, `read_video` | | **Search** | `search_files`, `find_files`, `file_info`, `get_tree`, `diff_files`, `get_diagnostics` | | **Shell** | `run_command`, `spawn_process`, `get_process_output`, `run_applescript` | | **Web Search** | `ddg_search`: free, tries DuckDuckGo, then Mojeek | | **Brave Search** | `web_search`, with a Brave Search API key | | **URL Fetch** | `fetch_url` | | **Git** | `git`: status, diff, log, blame, commit, branch, stash | | **Utilities** | `count_tokens`, `clipboard_read`, `clipboard_write`, `get_current_datetime`; `ask_user` is always on | **Max Tool Iterations** caps how many rounds of tool calls one message can take (10 by default). Shell and file tools act on real files, so point the working directory at a project, not your home folder, and use a model you trust with it. #### Use it from other apps: the local API Every model you start is served on your Mac. The **API Gateway** at http://localhost:8080 routes each request to the right model by the name in the request, and starts a stopped model when it's asked for. Each model also has its own port, from 8000 up. By default only your Mac can reach them; turning on **LAN** opens them to your network. **Servers & API** → **API** lists every endpoint for the models you're running, with copy-paste snippets in curl, Python and JavaScript. | Format | Endpoints | | --- | --- | | OpenAI | `/v1/chat/completions`, `/v1/responses`, `/v1/completions`, `/v1/models`, `/v1/embeddings` | | Anthropic | `/v1/messages` | | Ollama | `/api/chat`, `/api/generate`, `/api/tags`, `/api/show`, `/api/embed` | | Images | `/v1/images/generations`, `/v1/images/edits` | | Audio | `/v1/audio/transcriptions` (Whisper), `/v1/audio/speech` (Kokoro) | Behind the API, the engine caches what it has already read. With **Continuous Batching** on, the **Prefix Cache** lets a long chat or a coding agent's repeated instructions skip work on the next turn, **KV Cache Quantization** shrinks that cache, and **Disk Cache (Persistent)** keeps it on the SSD. These are in each model's server settings. A request in the OpenAI format, with the model name the API page shows: ``` curl http://localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "YOUR-MODEL", "messages": [{"role": "user", "content": "Hello!"}] }' ``` Any OpenAI SDK works with `base_url` set to http://localhost:8080/v1; the key can be any text unless you set one on the session. For coding agents, **Coding Tool Integration** sets up Claude Code, Codex CLI, OpenCode, OpenClaw and Hermes Agent with **Auto-Configure**, or shows a **Manual Setup**. For Claude Code, that means merging this into ~/.claude/settings.json: ``` { "env": { "ANTHROPIC_BASE_URL": "http://localhost:8080", "ANTHROPIC_MODEL": "YOUR-MODEL", "ANTHROPIC_API_KEY": "mlxharness" } } ``` #### The Network page, and MLX Node **Network** shows MLX Network live from mlx.app: Macs online and working, tokens per second, AI tokens today and countries. It checks whether MLX Node is installed on this Mac and running, and opens it or links to the download. Type your wallet name under **Your wallet** and press **Show** to watch its mined MLX count up; the name is kept on this Mac only. MLX Harness itself never mines. It runs models for you. [MLX Node](https://mlx.app/download) is the separate menu bar app that lends your Mac's idle time to MLX Network and earns MLX Tokens: up to 1 MLX a day on a free seat, locked for 30 days. If you use both, they share the same memory and GPU, so MLX Node's **Memory MLX may use** limit and manual Start and Stop keep it out of your way. MLX Tokens can lose value; read the [risks](https://mlx.app/risks). #### Questions **Is MLX Harness free?** Yes. It's free and licensed under the Apache License 2.0, like vMLX, the project it's built on. **When can I download MLX Harness?** Now. It's a free download at harness.mlx.app and mlx.app/download, for Apple Silicon Macs on macOS 14.5 or later. The source is on GitHub at github.com/MLXapp/harness. **Which Macs does MLX Harness run on?** Macs with Apple Silicon (M1 or later) and macOS 14.5 or later. Your Mac's memory decides which models fit; Compare on mlx.app shows them. **Does MLX Harness send my prompts anywhere?** No. Local models run on your Mac, and the API listens only on your Mac unless you turn on LAN. Prompts leave the Mac only if you connect a remote endpoint. **Does MLX Harness earn MLX Tokens?** No. MLX Harness runs models for you and only reads public network figures. MLX Node is the app that earns MLX Tokens for AI work. ### Mining crypto on a Mac URL: https://mlx.app/learn/mine-crypto-on-mac · Updated 2026-10-19 · Earn with your Mac **The short version** - Mining Bitcoin on a Mac earns, in practice, nothing. The network is run by purpose-built chips that are hundreds of thousands of times faster. - Coins mined on ordinary processors still exist, but on a Mac they rarely earn more than the electricity costs. - What Apple Silicon is good at is AI: lots of fast memory for running language models, at tens of watts rather than hundreds. - MLX Node uses exactly that. Your Mac runs AI work for the network and earns MLX Tokens, within the limits you set. Most guides to mining on a Mac skip the part that matters: what it actually earns. Here is the honest math for Apple Silicon, and the one kind of mining a Mac is genuinely good at. #### Can you mine Bitcoin on a Mac? Technically yes. Usefully, no. Bitcoin is mined by guessing hashes, and the whole network now makes about **1,000 exahashes a second**, a thousand billion billion guesses ([mempool.space](https://mempool.space/graphs/mining/hashrate-difficulty)). Almost all of that comes from ASICs, chips built to do nothing else. Be generous and say a Mac could manage a billion hashes a second. It would then be one part in about a trillion of the network. Mining alone, it would find a block about once every 19 million years. Joining a pool turns that into a steady trickle, but the trickle is far smaller than the electricity the Mac burns to make it. > Warning: Anything that promises to mine Bitcoin "on your Mac" at a profit is either paying you from something else or not telling you the truth. Treat it with care. #### Coins a Mac can mine Some coins are designed to be mined on ordinary processors rather than ASICs, Monero being the best known. A Mac can mine these, either with open-source miners or through mining marketplaces. The catch is the same: earnings move with the coin's price and the network's difficulty, and for a laptop or a Mac mini they are usually small next to the power cost. Before you start, put your Mac's numbers into a mining calculator for the coin and set them against your electricity price. Our guide to [how much power a Mac uses](https://mlx.app/learn/mac-power-consumption) has Apple's figures. Mining also keeps the processor at full load for hours. Apple Silicon handles that well, but a MacBook will run warm and its fans, if it has them, will be audible. #### Why Apple Silicon is good at something else Hashing rewards raw, repetitive arithmetic. AI rewards something Macs have in unusual amounts: fast memory that the processor and graphics share. To write each word, a language model has to read its weights from memory, so how fast it runs depends mostly on memory bandwidth ([llama.cpp benchmarks](https://github.com/ggml-org/llama.cpp/discussions/4167)). An M5 Max reads memory at up to 614 GB/s and holds up to 128 GB ([Apple](https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/)). That's enough to run a 120-billion-parameter model on a laptop. In crypto terms, a Mac is a poor hasher and a very good inference machine. So the question becomes: can a Mac earn by running AI instead of guessing hashes? That is what MLX Network was built for. #### Mining with AI: how MLX works **MLX Node** is a free menu bar app for Macs with Apple Silicon. Once it's linked to your wallet, your Mac takes AI jobs from the network, runs them with Apple's MLX framework and sends back the answers. Answers are checked against other Macs', and every hour the network pays MLX Tokens for verified work. 1. [Download MLX Node](https://mlx.app/download) and open it from the menu bar. 2. Create a passkey wallet or link the one you have. It takes a few seconds and needs no seed phrase to start. 3. Choose how your Mac works. New installs start in manual mode: once linked and plugged in, MLX Node works within your limits until you press Stop. Or let it work only after the Mac has been idle. 4. Watch it earn in the app, in your wallet and on [MLX World](https://world.mlx.app), the live map of the network. On a free seat, a Mac mines up to 1 MLX a day. A **dedicated seat** matched to the Mac's chip multiplies that: up to 4 MLX a day with a $50 Nano seat, 20 with a Core seat and 320 with an Ultra seat running on a Mac Studio. A seat works like a stake: it reserves your Mac's place on the network and sets its share of the rewards; paid AI jobs, when they arrive, are planned to go to Macs with seats first. Mined MLX is locked for 30 days from the day it is earned on a free seat, 3 to 20 with a dedicated seat. See [dedicated seats](https://mlx.app/nodes) for the details. #### Hash mining and AI mining, side by side | | Bitcoin on a Mac | CPU coins on a Mac | MLX Node | | --- | --- | --- | --- | | What the Mac does | Guesses hashes | Guesses hashes | Runs AI models | | Competes with | ASIC farms | Every CPU and botnet | Other Macs | | Uses the Mac's strengths | No | Partly | Yes: unified memory and bandwidth | | Earnings | Effectively zero | Usually below power cost | MLX Tokens, locked 30 days | | You decide when it runs | Yes | Yes | Yes, with battery and heat limits | MLX Tokens are a new token. Their price can go down as well as up, and they could become worthless. Read the [risks](https://mlx.app/risks) before you buy anything. #### Questions **Is it worth mining Bitcoin on a MacBook?** No. The Bitcoin network runs at about 1,000 exahashes a second, almost all of it on ASICs. A Mac's share is so small that its earnings, even in a pool, are far below the cost of the electricity it uses. **Will mining damage my Mac?** Apple Silicon is designed to throttle itself before it overheats, but hours at full load mean heat, fan noise and battery wear on laptops. MLX Node lets you set limits: it can pause on battery or when the Mac gets warm. **Does MLX Node mine crypto?** Not by hashing. It runs AI work for the network and earns MLX Tokens for verified results. It's free, you choose when it works, and mined MLX is locked for 30 days, or 3 to 20 with a dedicated seat. **Which Macs can run MLX Node?** Any Mac with Apple Silicon, from M1 onwards: MacBook Air, MacBook Pro, iMac, Mac mini, Mac Studio and Mac Pro. Intel Macs aren't supported. ### Earn with your Mac URL: https://mlx.app/learn/make-money-with-your-mac · Updated 2026-10-19 · Earn with your Mac **The short version** - There are four ways a Mac can earn on its own: sharing internet bandwidth, mining coins, renting out compute, and running AI work. - None of them is a salary. Earnings are small, they vary, and some are paid in tokens whose price moves. - Apple Silicon's real strength is AI: fast shared memory at low power. Work that uses it pays the Mac for what it is good at. - MLX Node is free, runs AI work only within the limits you set, and earns MLX Tokens. A Mac spends most of its life waiting. Several kinds of apps promise to turn that waiting into money. Here is what each one does, what it asks of your Mac, and what to expect. #### Four ways a Mac can earn ##### 1. Sharing your internet connection Bandwidth-sharing apps route other companies' web traffic through your connection, typically for data collection or ad verification, and pay you per gigabyte. They use almost no processor, so earnings depend on your connection and location, not on your Mac. Read what traffic passes through your IP address before you install one. ##### 2. Mining coins Your Mac guesses hashes for a cryptocurrency. For Bitcoin this is pointless on a Mac, and for coins mined on ordinary processors earnings are usually below the electricity cost. Our [guide to mining on a Mac](https://mlx.app/learn/mine-crypto-on-mac) has the numbers. ##### 3. Renting out compute Marketplaces rent your machine's processor or graphics card to people who need it. Most are built around Nvidia graphics cards and Windows or Linux, so Macs are rarely supported or rarely chosen. See [selling computing power](https://mlx.app/learn/sell-computing-power). ##### 4. Running AI work A network sends your Mac AI jobs, such as running a language model on a prompt, and pays for verified results. This is the one that fits Apple Silicon, because language models need exactly what Macs have: lots of fast memory shared by the processor and graphics. #### How MLX Node earns MLX Network pools Macs into one AI supercomputer. **MLX Node** lives in the menu bar. When it works, it runs AI jobs with Apple's MLX framework, other Macs check the answers, and every hour the network pays MLX Tokens for the verified work. - **Free.** The app costs nothing and never asks for payment to mine. - **On your terms.** New installs start in manual mode: once your Mac is linked and plugged in, MLX Node works within your limits until you press Stop. You can instead let it work only after the Mac has been idle, and it pauses when the Mac gets warm. - **Visible.** Every Mac appears on [MLX World](https://world.mlx.app), and every reward is on the public [explorer](https://explorer.mlx.app). - **Remote control.** MLX.app for iPhone and iPad shows your Macs, pauses or resumes them and tells you when one stops. On a free seat, a Mac mines up to 1 MLX a day. A dedicated seat matched to its chip multiplies that, from 4 MLX a day with a Nano seat to 320 with an Ultra seat. A seat works like a stake: it reserves your Mac's place on the network and sets its share of each hour's rewards. Mined MLX unlocks 30 days after the day it's earned on a free seat, and after the seat's own lock-in (3 to 20 days) with a dedicated seat. Paid jobs from outside clients, paid in USDC and never locked, are planned next, and are planned to go to Macs with dedicated seats first. #### What it costs to run The main cost is electricity. Apple Silicon is efficient: an M4 Mac mini draws about 4 W at idle and up to 65 W at full processor load ([Apple](https://support.apple.com/en-us/103253)). Even flat out around the clock, that is about 1.6 kWh a day, roughly 29 cents at the average US home price of 18.3 cents per kWh ([EIA, July 2026](https://www.eia.gov/electricity/monthly/update/end-use.php)). Real AI work usually draws less than the maximum. Laptops add battery wear if they work unplugged, which is why MLX can pause whenever the Mac is on battery. Our guide to [Mac power consumption](https://mlx.app/learn/mac-power-consumption) has the figures for every Mac mini and Mac Studio. #### What to expect, honestly Any app that pays your Mac is paying from somewhere: advertisers, buyers of compute, or a token. Ask which before you install. With MLX the payment is MLX Tokens from a fixed compute pool. Their value in dollars depends on what people pay for MLX, which can fall as well as rise. Earnings are not income you can count on, and you should never buy a dedicated seat with money you can't afford to lose. The [risk disclosure](https://mlx.app/risks) explains the rest. #### Questions **Can a Mac mini make money on its own?** Yes, a little, with apps that use its spare capacity. The option that suits a Mac mini best is AI work, because it has fast unified memory and low power draw. MLX Node is free and earns MLX Tokens for verified AI work. **Is it safe to leave my Mac working overnight?** Apple Silicon Macs are built to run under load and throttle before they overheat. With MLX you can set limits so it pauses on battery or when the Mac gets warm, and MLX.app on iPhone tells you if a Mac stops. **How much can I earn?** With MLX Node, up to 1 MLX a day on a free seat and up to 20 MLX a day with a Core seat, more with bigger seats. What that's worth in dollars depends on the MLX price, which moves. Nothing is guaranteed. **Do I need to buy anything?** No. MLX Node is free and mines on a free seat. A dedicated seat is optional; it multiplies what a Mac of its class earns. ### Sell your computing power URL: https://mlx.app/learn/sell-computing-power · Updated 2026-10-19 · Earn with your Mac **The short version** - Compute marketplaces pay you to run other people's workloads on your machine. - Most are built for Nvidia graphics cards on Windows or Linux, so Macs are rarely supported or chosen. - AI inference, running models on prompts, is the workload that suits Apple Silicon best. - MLX Network is built for Macs only: MLX Node runs AI work and earns MLX Tokens. AI companies spend fortunes on compute, and many people have powerful computers sitting idle. Here is how renting out spare compute actually works, and where a Mac fits in. #### Why compute is worth selling AI runs on compute, and there isn't enough of it. Alphabet, Amazon, Microsoft and Meta planned to spend more than $700 billion on capital expenditure in 2026, much of it for AI infrastructure ([Sherwood News](https://sherwood.news/tech/alphabet-amazon-microsoft-meta-plan-more-than-700-billion-on-capex-this-year/)). Data center electricity use grew sharply in 2025 ([IEA](https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions)). Meanwhile, hundreds of millions of capable computers sit idle most of the day. Spare-compute networks try to connect the two. #### How compute marketplaces work 1. You install a client that measures your hardware and offers it to the network. 2. Buyers send jobs: rendering, training, inference or batch processing. 3. Your machine runs the job in an isolated environment and returns the result. 4. The network checks the work and pays you, in dollars, credits or a token. The hard parts are trust and fit. Buyers need to know the work was done correctly, and the job has to suit your hardware. A job that needs an Nvidia graphics card with 24 GB won't run on a laptop, however fast. #### Why most marketplaces skip Macs Most AI software was written for Nvidia's CUDA, which Macs don't run. That's why most GPU rental platforms list Nvidia cards and little else. But Apple Silicon has an unusual advantage for AI inference: the processor and graphics share one large pool of fast memory. A Mac Studio with M5 Ultra holds up to 512 GB at 1.2 TB/s ([Apple](https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/)), enough to run models with hundreds of billions of parameters on one machine. Apple's own MLX framework is built to use it ([GitHub](https://github.com/ml-explore/mlx)). So Macs are good at AI work. What they've lacked is a network built for them. #### Selling Mac compute with MLX Network MLX Network is a compute network for Apple Silicon only. Work is placed by memory and bandwidth rather than raw speed, so every job runs whole on one Mac. Answers are checked against other Macs, and every hour the network pays MLX Tokens for the verified work. - Install [MLX Node](https://mlx.app/download), free, on any Mac with an M-series chip. - Link a passkey wallet. Your earnings and your Mac's work are on the public ledger. - Choose when it works, or let it work after the Mac is idle, with battery and heat limits. - Optionally add a [dedicated seat](https://mlx.app/nodes) matched to your Mac's chip to multiply what it earns. Today the network pays from its compute pool in MLX Tokens, locked for 30 days from the day they're earned (3 to 20 with a dedicated seat). Paid jobs from outside clients, paid in USDC and never locked, are planned next. The [whitepaper](https://mlx.app/whitepaper) explains how work is placed and verified. #### Questions **Can I rent out my Mac for AI?** Yes. Most GPU marketplaces are built for Nvidia hardware, but MLX Network is built for Apple Silicon Macs: MLX Node runs AI work for the network and earns MLX Tokens for verified results. **Will other people see my files?** No. MLX Node is a sandboxed app: it can reach the network and its own files, not your documents. Jobs are prompts for a model, and the answers go back to the network. **How is the work verified?** Each Mac's answers are compared with other Macs' answers to the same jobs, and the network also sends test jobs with known answers. Macs that disagree earn less or nothing. **Is selling compute passive income?** It's closer to a small, variable reward for lending your hardware. Earnings depend on how much your Mac works and on the value of what you're paid in. Nothing is guaranteed. ### How much power does a Mac use? URL: https://mlx.app/learn/mac-power-consumption · Updated 2026-10-19 · Earn with your Mac **The short version** - An M4 Mac mini uses about 4 W at idle and up to 65 W at full processor load. A Mac Studio with M5 Ultra uses 9 W at idle and up to 385 W. - Running an M4 Mac mini flat out for 24 hours costs about 29 cents at the average US home electricity price. - AI work rarely holds a Mac at its maximum, so real costs are usually lower. - An Apple Silicon Mac reaches high memory bandwidth, which is what AI output needs, at a fraction of a datacenter GPU's power. Apple Silicon Macs are some of the most efficient computers you can buy. Here are Apple's own figures, what a day of work costs, and what that means for running AI at home. #### Mac mini power consumption Apple's figures: idle with only the Finder open, maximum with a test that keeps the processor fully busy. Source: [Apple](https://support.apple.com/en-us/103253). | Mac mini | Idle | Maximum | | --- | --- | --- | | M1 (2020) | 6.8 W | 39 W | | M2 (2023) | 7 W | 50 W | | M2 Pro (2023) | 7 W | 100 W | | M4 (2024) | 4 W | 65 W | | M4 Pro (2024) | 5 W | 140 W | | M5 Pro (2026) | 6 W | 145 W | | M6 (2026) | 4 W | 70 W | #### Mac Studio power consumption Source: [Apple](https://support.apple.com/en-us/102027). Figures for laptops depend on the screen and battery charging, so Apple doesn't publish them the same way. | Mac Studio | Idle | Maximum | | --- | --- | --- | | M1 Max | 11 W | 115 W | | M1 Ultra | 13 W | 215 W | | M2 Max | 9 W | 145 W | | M2 Ultra | 10 W | 295 W | | M4 Max | 6 W | 145 W | | M3 Ultra | 9 W | 270 W | | M5 Max | 7 W | 200 W | | M5 Ultra | 9 W | 385 W | #### What a day of work costs Multiply watts by hours, divide by 1,000 for kilowatt-hours, and multiply by your electricity price. The average US home paid 18.3 cents per kWh in July 2026 ([EIA](https://www.eia.gov/electricity/monthly/update/end-use.php)); European prices are often higher ([Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Electricity_price_statistics)). Upper bounds: these assume the Mac never drops below its maximum, which real work rarely does. | 24 hours at maximum | Energy | At 18.3 ¢/kWh | | --- | --- | --- | | Mac mini, M4 (65 W) | 1.56 kWh | about $0.29 | | Mac mini, M4 Pro (140 W) | 3.36 kWh | about $0.61 | | Mac Studio, M5 Max (200 W) | 4.8 kWh | about $0.88 | | Mac Studio, M5 Ultra (385 W) | 9.24 kWh | about $1.69 | #### Why Macs are efficient for AI Writing text with a language model is limited by how fast weights can be read from memory, not by raw arithmetic ([llama.cpp benchmarks](https://github.com/ggml-org/llama.cpp/discussions/4167)). Apple Silicon puts fast memory right next to the processor and graphics, so a Mac reaches high memory bandwidth at a fraction of a datacenter GPU's power: Nvidia rates its H100 at up to 700 W ([Nvidia](https://www.nvidia.com/en-us/data-center/h100/)). That's why MLX Network is built on Macs. If you run [MLX Node](https://mlx.app/download), you can cap how hard it works, pause it on battery and let it rest when the Mac gets warm. On a laptop, keep it plugged in while it works to spare the battery. #### Questions **How many watts does a Mac mini use?** An M4 Mac mini uses about 4 W at idle and up to 65 W at full processor load; the M4 Pro model up to 140 W. Apple lists every model's figures on its support site. **Is it expensive to leave a Mac on all the time?** Not for an Apple Silicon Mac at idle: a few watts, a few dollars a year. Under constant full load, an M4 Mac mini costs about 29 cents a day at the average US electricity price. **Does running AI wear out my Mac?** Macs are built to run under load and slow themselves down before they overheat. The main wear item is a laptop's battery, so it's best to run sustained work on power. MLX Node can pause automatically on battery. ### DeepSeek V4.1 Flash 1017 URL: https://mlx.app/learn/deepseek-v4-1-flash-1017 · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - DeepSeek-V4.1-Flash-1017 is an update to DeepSeek-V4.1-Flash, released on 10 September 2026. Same architecture and API name, better scores and shorter answers. - In DeepSeek's own tests it gains most on agentic coding: 42.3 on Terminal-Bench 4.0, up from 31.2, and 77.0 on DeepSWE v1.1. - It is a mixture-of-experts model with 552 billion parameters, of which 16 billion work on each generated token, and a context of up to 1 million tokens. The weights are free under the MIT license. - The full weights take about 510 GB. On a Mac it needs a Mac Studio with 256 to 512 GB of memory, or SSD streaming on 128 GB. On 17 October 2026 DeepSeek upgraded V4.1 Flash, its open-weight model, to a version named 1017. Here's what changed, how it compares, what it costs through the API, and what it takes to run it yourself. #### What DeepSeek V4.1 Flash 1017 is DeepSeek-V4.1-Flash is the smallest model of DeepSeek's V4.1 family, with built-in image understanding. It arrived in DeepSeek's API on 10 September 2026; on 17 October DeepSeek replaced it with DeepSeek-V4.1-Flash-1017, which it calls a minor update ([DeepSeek API changelog](https://api-docs.deepseek.com/updates/)). The update improves reasoning, coding, agentic and visual benchmarks, and the model finishes tasks with shorter output in less time. DeepSeek tuned it for its own DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness 0.2.1-rc.1 or later ([Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017)). The name follows DeepSeek's habit of dating updates: 1017 is the month and day, 17 October. #### Benchmarks: 1017 against V4.1 Flash and others DeepSeek's own figures from the [model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017); GLM-5.3's HLE score is on the text-only subset. Independent results had not been published when this guide was written. | Benchmark | V4.1 Flash 1017 | V4.1 Flash | Opus 5 | GPT-5.6 Sol | Kimi K3 | GLM-5.3 | | --- | --- | --- | --- | --- | --- | --- | | HLE, no tools | 42.6 | 36.8 | 56.3 | 44.5 | 43.5 | 42.0 | | Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 | 39.9 | 12.6 | 37.9 | | DeepSWE v1.1 | 77.0 | 74.2 | 74.0 | 73.0 | 67.5 | 66.9 | | ProgramBench | 29.1 | 20.3 | 37.0 | 23.0 | 17.5 | 19.0 | | NL2Repo-Bench | 68.2 | 65.4 | 75.3 | 56.8 | 58.0 | 58.0 | | Agents' Last Exam | 35.4 | 31.8 | 28.6 | 26.7 | 27.6 | 28.5 | | Apex Bench v1.1 | 72.1 | 64.6 | 65.8 | 54.3 | 47.8 | 56.6 | On these numbers, 1017 leads every model in the table on DeepSWE, Agents' Last Exam and Apex Bench, and trails Opus 5 on reasoning without tools, terminal work and building whole programs. It also improves on chart and image tests such as Chartography (83.0, up from 78.9) and BabyVision (92.5, up from 89.6). Cost per task falls too. In DeepSeek's charts reported by [The Decoder](https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/), a DeepSWE task costs just under $0.20 at the highest setting, against about $0.50 for the previous version. #### Parameters, context and size From DeepSeek's model cards for [V4.1 Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) and [1017](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017). | | DeepSeek-V4.1-Flash-1017 | | --- | --- | | Total parameters | 552 billion (mixture of experts) | | Active per token | 8 billion reading the prompt, 16 billion writing | | Experts | 384 routed and 1 shared, 6 used per token | | Context | Up to 1,048,576 tokens | | Memory per token of context | About 890 bytes, a quarter of V4 Flash | | Weights | FP8, with experts in FP4; about 510 GB | | Input | Text and images | | License | MIT | About 189 GiB of the download is two Engram tables: a large memory the model looks up rather than computes, a few rows per token. Runtimes such as DwarfStar keep them on the SSD and read only the rows they need, which is why the model can run in far less memory than its download size ([antirez on Hugging Face](https://huggingface.co/antirez/deepseek-v4.1-flash-gguf)). #### API name and price In DeepSeek's API the model is called `deepseek-flash`, and it now points to 1017; the older names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are routed to it for now. DeepSeek says the calling method and prices stayed the same with 1017 ([DeepSeek](https://api-docs.deepseek.com/updates/)). Prices listed on [OpenRouter](https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017) on 19 October 2026. Cached input costs a fraction of a cent per million tokens. Prices change; check before you build on them. | Provider | Input, per million tokens | Output, per million tokens | | --- | --- | --- | | DeepSeek | $0.15 | $0.60 | | Cloudflare | $0.10 | $0.60 | | DeepInfra | $0.098 | $0.294 | | Together, Fireworks, Baseten | $0.30 | $1.20 | DeepSeek recommends a temperature of 1.0, top_p of 1.0 and room for long answers (max_tokens of at least 256K). The weights ship without a Jinja chat template: prompts are built with the encoding code in the model's repository. #### Can you run it on your own Mac? Yes, but only on the biggest Macs. A Mac Studio with 512 GB of memory (M3 Ultra or M5 Ultra) holds the 4-bit builds entirely; 256 GB is enough for a 2-bit build with the Engram tables left on the SSD; a 128 GB Mac can stream from the SSD, slowly. The first published run, on an M3 Ultra with 512 GB, generated about 6.5 tokens a second. Which build to download, the commands and the speeds are in our step-by-step guide: [Run DeepSeek V4.1 Flash on a Mac](https://mlx.app/learn/run-deepseek-v4-1-flash-on-mac). For every other Mac, the API costs cents per million tokens, and smaller open models run well locally: see [the best Mac for local LLMs](https://mlx.app/learn/best-mac-for-local-llm). #### Put a big Mac's idle hours to work A Mac Studio bought to run models like this spends most of the day waiting. [MLX Node](https://mlx.app/download) lends that spare time to MLX Network, which runs AI jobs on Macs with Apple's MLX framework, and earns MLX Tokens for verified work: up to 1 MLX a day on a free seat, and up to 320 with an Ultra [dedicated seat](https://mlx.app/nodes) running on an Ultra Mac. Mined MLX is locked for 30 days, less with a dedicated seat; see the [risks](https://mlx.app/risks). #### Questions **What is DeepSeek V4.1 Flash 1017?** It is the 17 October 2026 update of DeepSeek-V4.1-Flash, DeepSeek's open-weight mixture-of-experts model. It keeps the architecture and API name and scores higher on reasoning, coding, agent and image benchmarks. **What is the difference between V4.1 Flash and 1017?** Same model size and API, newer training. In DeepSeek's tests 1017 scores 42.3 instead of 31.2 on Terminal-Bench 4.0 and 77.0 instead of 74.2 on DeepSWE v1.1, and it writes shorter answers, so tasks cost less. **How many parameters does DeepSeek V4.1 Flash have?** 552 billion in total. It activates 8 billion per token while reading the prompt and 16 billion while writing, and adds two large Engram lookup tables that can stay on disk. **Is DeepSeek V4.1 Flash 1017 free?** The weights are free to download and use under the MIT license. Using it through DeepSeek's API is paid: $0.15 per million input tokens and $0.60 per million output tokens, as listed on OpenRouter. **Can I run DeepSeek V4.1 Flash 1017 locally?** Yes, with a lot of memory. On a Mac, a Mac Studio with 512 GB runs the 4-bit builds and 256 GB runs a 2-bit build with the Engram tables on the SSD. Smaller Macs are better off with the API or a smaller model. ### Run DeepSeek V4.1 Flash on a Mac URL: https://mlx.app/learn/run-deepseek-v4-1-flash-on-mac · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - You need a Mac Studio with 512 GB for the 4-bit and MLX builds, or 256 GB for a 2-bit build that leaves the Engram tables on the SSD. - A 128 GB MacBook Pro or Mac Studio can run the 2-bit build by streaming from the SSD, or split it with a second 128 GB Mac over Thunderbolt 5. - The first published Mac run, on an M3 Ultra with 512 GB, wrote about 6.5 tokens a second and read prompts at about 275 tokens a second. - Three ways to run it: DwarfStar (ds4), a llama.cpp branch for the GGUF files, or MLX builds through mlx-vlm and LM Studio. DeepSeek V4.1 Flash is open and one of the strongest models you can download, but it is big: about 510 GB of weights. Here's which Macs can hold it, which build to pick, how fast it runs and the commands to start. #### How much memory it needs The model has 552 billion parameters, but about 189 GiB of its download is two Engram lookup tables that are read a few rows per token. Runtimes that leave those tables on the SSD need far less memory than the download size; builds that load everything need it all ([antirez on Hugging Face](https://huggingface.co/antirez/deepseek-v4.1-flash-gguf)). Sizes from the builds' pages on Hugging Face: [DwarfStar](https://huggingface.co/antirez/deepseek-v4.1-flash-gguf), [GGUF 1017](https://huggingface.co/mrfakename/DeepSeek-V4.1-Flash-1017-GGUF), [MLX oQ4e](https://huggingface.co/mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-mtp), [MLX oQ2.5e](https://huggingface.co/mlx-community/DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp). Context and runtime buffers come on top. | Build | Weights | Download | Held in memory | Runs with | | --- | --- | --- | --- | --- | | DwarfStar Q2 | September | 366 GB | 163 GB, Engram on SSD | ds4 | | DwarfStar Q4 | September | 519 GB | 316 GB, Engram on SSD | ds4 | | GGUF Q2_K | 1017 | 265 GB | Not published | llama.cpp branch | | GGUF Q4_K_M | 1017 | 445 GB | Not published | llama.cpp branch | | MLX oQ2.5e | 1017 | 343 GB | The whole model | mlx-vlm, LM Studio | | MLX oQ4e | 1017 | 432 GB | The whole model | mlx-vlm, LM Studio | Leave room for macOS and for the context: at about 890 bytes per token, a 100,000-token conversation adds under 0.1 GB, so the weights are what decide. #### Which Mac can run it From the [DwarfStar](https://github.com/antirez/ds4) and build pages. The 512 GB Mac Studio is Apple's [M3 Ultra](https://support.apple.com/en-us/122211) or [M5 Ultra](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/) model. | Mac | What runs | How | | --- | --- | --- | | Mac Studio, 512 GB (M3 Ultra or M5 Ultra) | Every build, including 4-bit and the MLX 1017 builds | Fully in memory | | Mac Studio, 256 GB | DwarfStar Q2 | In memory, Engram tables on the SSD | | Two 128 GB Macs over Thunderbolt 5 | DwarfStar Q2 | Split between them with RDMA | | MacBook Pro or Mac Studio, 128 GB | DwarfStar Q2 | Streamed from the SSD, experimental and slower | | 64 GB or less | Not practical | Use the API or a smaller model | Keep the model on the Mac's internal SSD: Engram rows and streamed weights are read from disk during generation, and an external drive is much slower. Plan for the download too: 250 to 520 GB, plus free space while files are joined. #### How fast it runs The first public benchmark on a Mac ran a mixed 4- and 8-bit MLX build of the September weights on an M3 Ultra Mac Studio with 512 GB ([benchmark on GitHub](https://github.com/LocalAiCherry/deepseek-v41-mac-bench)): MLX 0.32.2 with a development version of mlx-lm, tested on 10 September 2026. | Prompt length | Reading the prompt | Writing | Peak memory | | --- | --- | --- | --- | | 8,192 tokens | 275 tokens/s | 6.5 tokens/s | 429 GB | | 32,768 tokens | 295 tokens/s | 6.3 tokens/s | 431 GB | That is far below what the hardware allows. Writing reads about 16 billion parameters per token, about 9 GB at 4 bits, so memory bandwidth caps one chat at roughly 90 tokens a second on an M3 Ultra (819 GB/s) and 130 on an M5 Ultra (1.2 TB/s). Support for the new architecture is weeks old, so expect newer runtimes to close part of the gap; compare current numbers on [MLX Bench](https://mlx.app/bench) and in the runtimes' release notes. Reading prompts is limited by compute rather than memory. M5-generation chips have neural accelerators in the GPU and read prompts several times faster than earlier chips with MLX ([Apple](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)). #### How to run it ##### DwarfStar (ds4): the leanest on memory DwarfStar is an open-source engine built for DeepSeek Flash models, with Metal support for Macs from 96 GB and SSD streaming for smaller ones. It keeps the Engram tables on disk in every mode ([GitHub](https://github.com/antirez/ds4)). ``` git clone https://github.com/antirez/ds4.git cd ds4 && make ./download_model.sh ds41f-q2 ./ds4 -m gguf/DeepSeek-V4.1-Flash-Q2.gguf --ssd-streaming --ctx 32768 ``` Leave out `--ssd-streaming` when the Mac has 256 GB or more. `./ds4-server` serves the model over a local API. ##### llama.cpp: the 1017 GGUF files Mainline llama.cpp and Ollama don't support V4.1 Flash yet. The 1017 GGUF files were made with a llama.cpp branch, which you build yourself; keeping the Engram tables on the CPU side is recommended ([Hugging Face](https://huggingface.co/mrfakename/DeepSeek-V4.1-Flash-1017-GGUF)). The files are text-only for now. ``` git clone https://github.com/vcruz305/llama.cpp && cd llama.cpp git checkout 2076412 cmake -B build && cmake --build build -j ./build/bin/llama-server -hf mrfakename/DeepSeek-V4.1-Flash-1017-GGUF:Q2_K -ot "engram_embd.weight=CPU" ``` ##### MLX: the 1017 builds for Apple Silicon The MLX community builds of 1017 keep DeepSeek's original FP4 and FP8 precision for most weights and include the multi-token prediction weights. They load through mlx-vlm, which also handles images, and LM Studio lists them too. They need a 512 GB Mac. ``` pip install -U mlx-vlm python -m mlx_vlm.generate --model mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-mtp --prompt "Hello" --max-tokens 200 ``` New to local models? Start with [running an LLM on a Mac](https://mlx.app/learn/run-llm-on-mac) and [MLX vs Ollama vs llama.cpp](https://mlx.app/learn/mlx-vs-ollama). #### No 512 GB Mac? Your options - **Use the API.** DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens for `deepseek-flash`; see [DeepSeek V4.1 Flash 1017](https://mlx.app/learn/deepseek-v4-1-flash-1017). - **Run a smaller open model.** A 64 GB Mac runs 70B-class models and a 128 GB Mac runs gpt-oss-120b well: see [the best Mac for local LLMs](https://mlx.app/learn/best-mac-for-local-llm). - **Join Macs together.** Two or more Macs over Thunderbolt 5 can hold models no single one can: see [Mac AI clusters](https://mlx.app/learn/mac-ai-cluster). - **Let your Mac earn.** [MLX Node](https://mlx.app/download) runs AI jobs for MLX Network on any Apple Silicon Mac and earns MLX Tokens, up to 1 MLX a day without a [dedicated seat](https://mlx.app/nodes). Mined MLX is locked for 30 days, less with a dedicated seat; see the [risks](https://mlx.app/risks). #### Questions **How much RAM do I need for DeepSeek V4.1 Flash?** About 163 GB for the 2-bit DwarfStar build with the Engram tables left on the SSD, about 316 GB for its 4-bit build, and the whole download, 343 to 432 GB, for the MLX builds. Add room for macOS and the context. **Can a MacBook Pro run DeepSeek V4.1 Flash?** Only a 128 GB one, and only by streaming the 2-bit build from the SSD with DwarfStar, which is experimental and slow. It runs fully in memory on a Mac Studio with 256 or 512 GB. **How fast is DeepSeek V4.1 Flash on an M3 Ultra?** The first public run on a 512 GB M3 Ultra wrote about 6.5 tokens a second and read prompts at about 275 tokens a second, with an MLX build. The memory bandwidth would allow about 90 tokens a second, so newer software should be faster. **Is there an MLX version of DeepSeek V4.1 Flash 1017?** Yes. The MLX community published 1017 builds of 343 GB and 432 GB. They run with mlx-vlm or LM Studio on a Mac with 512 GB of memory. **Does DeepSeek V4.1 Flash work in Ollama?** Not in standard Ollama or llama.cpp yet. The GGUF files need a llama.cpp branch built from source, or you can use DwarfStar, which supports the model on Macs. ### Run an LLM on your Mac URL: https://mlx.app/learn/run-llm-on-mac · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - Any Mac with Apple Silicon can run a local LLM. Memory decides which models fit; memory bandwidth decides how fast they run. - The quickest way to start: LM Studio (an app) or Ollama (one command). For the most speed and control: Apple's MLX with mlx-lm. - Rule of thumb: a 4-bit model needs a little over half a gigabyte of memory per billion parameters, plus room for the context. - Measure your Mac on MLX Bench and compare it with every other Apple Silicon chip. Apple Silicon Macs are some of the best machines for running language models locally: private, offline and fast. Here is how to start in five minutes, and how to pick a model that fits. #### What fits on your Mac A model's weights have to fit in memory with room to spare for the context and macOS. At 4 bits per weight, plan on about 0.55 to 0.6 GB per billion parameters. Approximate; quantization, context length and other apps change the numbers. Parameter counts from the model cards ([gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b), [Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)). | Mac memory | Comfortable model size (4-bit) | Examples of the class | | --- | --- | --- | | 8 GB | up to about 4B | small chat and coding models | | 16 GB | up to about 14B | most popular 7B–14B models | | 32 GB | up to about 30B, or mixture-of-experts models of similar size | mid-size reasoning and coding models | | 64 GB | up to about 70B | large dense models | | 128 GB | up to about 120B (mixture-of-experts) | gpt-oss-120b, about 65 GB | | 512 GB | hundreds of billions of parameters | Qwen3.5-397B-A17B, about 220 GB | Speed is a separate question: each new token needs the active weights read from memory once, so tokens per second scale with memory bandwidth. Our guide to the [best Mac for local LLMs](https://mlx.app/learn/best-mac-for-local-llm) covers every chip. #### The easy way: LM Studio [LM Studio](https://lmstudio.ai) is a free Mac app with a chat window and a model browser. On Apple Silicon it can run models in Apple's MLX format as well as the GGUF format used by llama.cpp. 1. Download LM Studio and open it. 2. Search for a model, prefer an MLX version on Apple Silicon, and pick a size that fits your memory. 3. Load it and start chatting. You can also turn on its local server for other apps. #### One command: Ollama [Ollama](https://ollama.com) runs models from the terminal and serves them to other apps on your Mac. After installing it, one command downloads and starts a model: ``` ollama run llama3.2 ``` Since version 0.19, Ollama on Apple Silicon has a preview powered by Apple's MLX framework for supported models, with large speedups; the preview asks for a Mac with more than 32 GB of memory ([Ollama](https://ollama.com/blog/mlx)). #### The fastest way: Apple's MLX and mlx-lm [MLX](https://github.com/ml-explore/mlx) is Apple's open-source machine learning framework for Apple Silicon, and **mlx-lm** is its package for language models. Thousands of models are ready to use from the mlx-community organization on Hugging Face. ``` pip install mlx-lm mlx_lm.generate --model mlx-community/Llama-3.2-3B-Instruct-4bit --prompt "Explain unified memory in one paragraph" ``` Use `mlx_lm.chat` for a conversation in the terminal and `mlx_lm.server` for a local API that other apps can call. On M5 chips, MLX uses the new neural accelerators in the GPU: Apple measured the first token arriving 3.3 to 4.1 times sooner than on M4 ([Apple Machine Learning Research](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)). #### Measure it, then put it to work [MLX Bench](https://mlx.app/bench) is an open leaderboard of MLX speed on every Apple Silicon chip: generation and prompt tokens per second and peak memory. Run the one-file script on your Mac and see where it lands. When you're not using your Mac, it can run AI for others. [MLX Node](https://mlx.app/download) is a free app that runs AI jobs for MLX Network on the same MLX framework and earns MLX Tokens for verified work, within the limits you set. #### Questions **Can a MacBook Air run an LLM?** Yes. A MacBook Air with 16 GB runs 7B to 14B models at 4-bit comfortably; with 32 GB, models up to about 30B. It will be slower than a Mac with a Pro or Max chip, which has more memory bandwidth. **Is MLX faster than Ollama on a Mac?** Often, because MLX is written for Apple Silicon. Ollama's own MLX-powered preview reports large speedups on Apple Silicon. Our MLX vs Ollama guide compares them in detail. **Do local models work offline?** Yes. Once downloaded, the model runs entirely on your Mac and your prompts never leave it. **How much memory do I need for a 70B model?** About 40 GB for the weights at 4-bit, so a Mac with 64 GB or more. Mixture-of-experts models with similar quality may need more memory but run faster. ### Download and run LLMs on your Mac URL: https://mlx.app/learn/download-and-run-llms-on-mac · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - Open models live on Hugging Face. On a Mac, look for MLX builds (often from mlx-community) or GGUF files for llama.cpp, Ollama and LM Studio. - Memory decides what fits: a 4-bit MLX model needs about 0.56 GB per billion parameters, plus room for context and macOS. - Gated models need a free Hugging Face account, an accepted licence and a token. - Run them with MLX Harness, mlx-lm, LM Studio or Ollama. Running a model locally starts with finding the right file: the right format for your app, a size your Mac's memory holds and a quantization that keeps it sharp. Here's how to choose, download and run open models on Apple Silicon. #### Where to find models [Hugging Face](https://huggingface.co/models) hosts most open-weight models. The same model usually comes in several builds, and the format decides which app can run it: Ready MLX builds of popular models are published by the [mlx-community](https://huggingface.co/mlx-community) organization; mlx-lm can make your own ([mlx-lm](https://github.com/ml-explore/mlx-lm)). | Format | What it looks like | Runs in | | --- | --- | --- | | MLX | A folder of safetensors files with config.json, names like Model-4bit | mlx-lm, LM Studio, MLX Harness | | GGUF | One or a few .gguf files, names like Q4_K_M | llama.cpp, Ollama, LM Studio | | Original weights | Full-precision safetensors from the model's maker | Convert to MLX first, or pick a ready build | Not sure which runtime to use? [MLX vs Ollama vs llama.cpp](https://mlx.app/learn/mlx-vs-ollama) compares them. New to all of this? Start with [running an LLM on a Mac](https://mlx.app/learn/run-llm-on-mac). #### Pick a size that fits your memory A model's weights have to fit in memory with room to spare. MLX stores a quantized weight as its bits plus a shared scale and bias for every group of 64 weights, about half a bit more per weight ([MLX](https://ml-explore.github.io/mlx/build/html/python/_autosummary/mlx.core.quantize.html)). So 4-bit costs about 0.56 GB per billion parameters and 8-bit about 1.06 GB. Weights only: parameters × (bits + 0.5) ÷ 8 for MLX quantized builds, 2 bytes per parameter at 16-bit. The conversation's context and macOS need room on top. | Parameters | 4-bit (MLX) | 8-bit (MLX) | 16-bit | | --- | --- | --- | --- | | 3 billion | about 1.7 GB | about 3.2 GB | 6 GB | | 8 billion | about 4.5 GB | about 8.5 GB | 16 GB | | 14 billion | about 7.9 GB | about 14.9 GB | 28 GB | | 32 billion | about 18 GB | about 34 GB | 64 GB | | 70 billion | about 39 GB | about 74 GB | 140 GB | macOS lets the GPU use only part of unified memory by default; [Compare](https://mlx.app/compare) assumes about three quarters, and mlx-lm's README shows how to raise the limit on macOS 15 or later with `sudo sysctl iogpu.wired_limit_mb=N` ([mlx-lm](https://github.com/ml-explore/mlx-lm)). For mixture-of-experts models, total parameters decide memory and active parameters decide speed. [Compare](https://mlx.app/compare) does this arithmetic for every popular model and every Apple chip, and [the best Mac for local LLMs](https://mlx.app/learn/best-mac-for-local-llm) explains why bandwidth sets the speed. #### Quantization: 4-bit, 8-bit and the rest Quantization stores each weight in fewer bits, so the model takes less memory and runs faster, at some cost in quality. 4-bit is the usual starting point on a Mac. 8-bit is closer to the original and twice the size. Below 4 bits, files get small but answers get worse sooner, so try them only when a bigger model wouldn't fit otherwise. - **MLX builds** say it in the name: -4bit, -8bit, and sometimes mixed-precision builds that keep sensitive layers at more bits. - **GGUF files** use llama.cpp's names: Q4_K_M, Q5_K_M, Q8_0 and so on. When you run a Hugging Face GGUF repository in Ollama without naming one, it picks Q4_K_M if it's there ([Hugging Face](https://huggingface.co/docs/hub/ollama)). - **Make your own** MLX build from original weights: `mlx_lm.convert --model mistralai/Mistral-7B-Instruct-v0.3 -q` writes a 4-bit copy ([mlx-lm](https://github.com/ml-explore/mlx-lm)). #### Disk space and where models are kept A download is about as big as the weights in the table above, so a 4-bit 8B model is a few gigabytes and a 70B model tens of gigabytes. Trying several sizes or builds of the same model adds up quickly. mlx-lm and other Hugging Face tools keep downloads in ~/.cache/huggingface/hub; set `HF_HOME` or `HF_HUB_CACHE` to move it. `hf cache ls` shows what's there and `hf cache rm` deletes a model ([Hugging Face](https://huggingface.co/docs/huggingface_hub/guides/manage-cache)). Ollama and LM Studio keep models in their own folders and manage them themselves. #### Gated models and Hugging Face tokens Some models, Llama among them, are gated: you need to accept the maker's terms before you can download them ([Hugging Face](https://huggingface.co/docs/hub/models-gated)). 1. Sign in to a free Hugging Face account and open the model's page. 2. Accept the terms with the button on the page (by default **Agree and send request to access repo**). Some makers approve requests by hand, which takes longer. 3. Create an access token in your Hugging Face settings. 4. Give the token to your tools: run `hf auth login` in Terminal for mlx-lm and other Hugging Face tools ([Hugging Face](https://huggingface.co/docs/huggingface_hub/quick-start#login)), or paste it into the app's Hugging Face token field. #### Run it: four ways ##### mlx-lm: Apple's MLX in Terminal Install it with pip, then name any MLX model on Hugging Face; the first run downloads it ([mlx-lm](https://github.com/ml-explore/mlx-lm)). ``` pip install mlx-lm mlx_lm.generate --model mlx-community/Llama-3.2-3B-Instruct-4bit --prompt "How tall is Mt Everest?" mlx_lm.chat --model mlx-community/Llama-3.2-3B-Instruct-4bit ``` `mlx_lm.server` serves a model on an OpenAI-style API at localhost:8080 for other apps ([mlx-lm](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md)): ``` mlx_lm.server --model mlx-community/Llama-3.2-3B-Instruct-4bit curl localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"messages": [{"role": "user", "content": "Say this is a test!"}]}' ``` ##### LM Studio: an app with a model browser [LM Studio](https://lmstudio.ai/docs/app) searches and downloads models from Hugging Face inside the app and runs MLX models on Apple Silicon as well as GGUF through llama.cpp. It needs Apple Silicon and macOS 14 or later, and recommends 16 GB of memory or more ([LM Studio](https://lmstudio.ai/docs/system-requirements)). ##### Ollama: one command [Ollama](https://ollama.com) runs models from its own library, and any GGUF repository on Hugging Face directly, optionally with a quantization tag ([Hugging Face](https://huggingface.co/docs/hub/ollama)): ``` ollama run hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:IQ3_M ``` ##### MLX Harness: our free app [MLX Harness](https://mlx.app/learn/how-to-use-mlx-harness) is our free Mac app, built on vMLX: it searches Hugging Face for MLX models, chats, runs tools and image models, and serves everything on a local API. Its gateway also uses port 8080, so don't run it alongside `mlx_lm.server` on the same port. When you're not using the Mac, [MLX Node](https://mlx.app/download) can lend it to MLX Network instead. #### Questions **Where do I download LLMs for a Mac?** From Hugging Face. Look for MLX builds, many published by mlx-community, or GGUF files. LM Studio and Ollama can search and download for you. **Which model size should I download?** One whose weights fit comfortably in your Mac's memory: at 4-bit about 0.56 GB per billion parameters, so about 4.5 GB for an 8B model and 39 GB for 70B, plus room for context. Compare on mlx.app checks it for your Mac. **MLX or GGUF on a Mac?** Both run well on Apple Silicon. MLX builds are made for Apple's MLX and run in mlx-lm, LM Studio and MLX Harness; GGUF files run in llama.cpp, Ollama and LM Studio. **Do I need a Hugging Face account?** Not for public models. Gated models need a free account, an accepted licence and an access token. **Can I run downloaded models offline?** Yes. Once the files are on your Mac, the model runs without an internet connection and your prompts stay on the Mac. ### The best Mac for local LLMs URL: https://mlx.app/learn/best-mac-for-local-llm · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - Memory decides which models fit. Buy as much as you can; it can't be upgraded later. - Memory bandwidth decides how many tokens a second you get. Max and Ultra chips have three to twelve times the bandwidth of the base chip of their generation. - Best value for most people: a Mac mini or MacBook Pro with a Pro chip and 64 GB. For the biggest models: Mac Studio with Max or Ultra. - Check real numbers on MLX Bench before you buy. For running language models, two numbers matter more than anything else on the spec sheet: how much unified memory the Mac has, and how fast it can read it. Here's how every Apple Silicon chip compares. #### Why memory and bandwidth decide everything A language model writes one token at a time, and for each one it reads every active weight from memory. So the speed limit for a single conversation is roughly memory bandwidth divided by the size of the active weights. In llama.cpp's public benchmarks, real generation reaches 43% to 80% of that limit on Apple chips ([llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/4167)). Reading your prompt is different: that part is limited by compute, which is why M5-generation chips, with neural accelerators in the GPU, process prompts several times faster than earlier ones ([Apple](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)). #### Every chip, compared Apple's figures for the top configuration of each chip: [M4 Pro and Max](https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/), [M5](https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/), [M5 Pro and Max](https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/), [M6 and M5 Ultra](https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/), [M3 Ultra](https://support.apple.com/en-us/122211). Lower configurations can have less. | Chip | Max memory | Bandwidth | Found in | | --- | --- | --- | --- | | M5 | 32 GB | 153 GB/s | MacBook Air, 14-inch MacBook Pro | | M6 | 32 GB | 170 GB/s | Mac mini | | M4 Pro | 64 GB | 273 GB/s | Mac mini, MacBook Pro | | M5 Pro | 64 GB | 307 GB/s | MacBook Pro, Mac mini | | M4 Max | 128 GB | 546 GB/s | MacBook Pro, Mac Studio | | M5 Max | 128 GB | 614 GB/s | MacBook Pro, Mac Studio | | M3 Ultra | 512 GB | 819 GB/s | Mac Studio (2025) | | M5 Ultra | 512 GB | 1.2 TB/s | Mac Studio (2026) | Older chips still work well when they have the memory. An M1 Max or M2 Max with 64 GB reads at 400 GB/s, more than any base chip today, and often costs far less second-hand. #### Which Mac to buy ##### Just trying it: MacBook Air or Mac mini with 16 to 32 GB Runs 7B to 14B models comfortably and 30B-class models at 32 GB. Fine for chat, writing and light coding help. ##### Best value: Pro chip with 64 GB A Mac mini or MacBook Pro with M4 Pro or M5 Pro holds 70B-class models and runs mid-size models at a good pace. The sweet spot for most developers. ##### Serious local AI: Max chip with 128 GB Runs gpt-oss-120b (about 65 GB at 4-bit) on a laptop; from Apple's bandwidth figure, the theoretical ceiling for one stream on M5 Max is about 225 tokens a second. The best portable choice. ##### The biggest models: Mac Studio with Ultra Up to 512 GB at 1.2 TB/s with M5 Ultra holds models with hundreds of billions of parameters on one desk. For even bigger models, Macs can be joined into a cluster over Thunderbolt 5: see [Mac AI clusters](https://mlx.app/learn/mac-ai-cluster). #### Make the Mac pay its way A Mac bought for AI spends most of its time waiting. [MLX Node](https://mlx.app/download) is a free app that lends that spare capacity to MLX Network, running AI jobs on the same MLX framework and earning MLX Tokens for verified work. A [dedicated seat](https://mlx.app/nodes) matched to the chip multiplies what it earns: a Pro seat runs at full rate on a Pro or Max chip, an Ultra seat on an Ultra chip. Compare your Mac on [MLX Bench](https://mlx.app/bench) first: it shows measured tokens per second for open models on every chip. #### Questions **Is 16 GB enough for a local LLM?** Yes for models up to about 14B parameters at 4-bit. For larger models, or long contexts, 32 GB or more is much more comfortable. **Mac Studio or MacBook Pro for LLMs?** Both come with Max chips and up to 128 GB. The MacBook Pro is portable; the Mac Studio stays cooler under long loads and is the only Mac with Ultra chips and up to 512 GB. **Is a Mac better than an Nvidia GPU for local AI?** For the largest models at home, often yes: a Mac's unified memory can hold far more than a consumer graphics card's memory. A high-end Nvidia card is faster on models that fit in its memory. **Does the M5 make LLMs faster?** Yes, mainly for reading prompts: Apple measured the first token arriving 3.3 to 4.1 times sooner on M5 than M4 with MLX. Generation speed grows with memory bandwidth, which rose 28% from M4 to M5. ### Mac AI clusters URL: https://mlx.app/learn/mac-ai-cluster · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - macOS 26.2 added RDMA over Thunderbolt 5, letting Macs read each other's memory with very low latency. - Apple's MLX framework and the open-source exo project both split one model across several Macs. - Four M3 Ultra Mac Studios ran a one-trillion-parameter model at about 30 tokens a second in independent tests. - A cluster is a few Macs on one desk. MLX Network connects Macs everywhere. When a model is too big for one Mac, you can join several. Since macOS 26.2, Macs can share memory directly over Thunderbolt 5. Here's how Mac clusters work, what they deliver, and where a network of Macs fits in. #### What changed: RDMA over Thunderbolt 5 Since macOS 26.2, Macs can exchange memory directly over Thunderbolt 5 using RDMA, a feature Apple names for distributed AI inference with MLX ([Apple](https://developer.apple.com/documentation/macos-release-notes/macos-26_2-release-notes)). MLX exposes it through its JACCL backend, next to MPI and TCP rings ([MLX docs](https://ml-explore.github.io/mlx/build/html/usage/distributed.html)). The gain is latency: RDMA lowers memory-access latency from about 300 µs to under 50 µs, according to MLX's developers. In Jeff Geerling's tests, four M3 Ultra Mac Studios ran the 1-trillion-parameter Kimi K2 Thinking at about 30 tokens a second ([Jeff Geerling](https://jeffgeerling.com/blog/2025/15-tb-vram-on-mac-studio-rdma-over-thunderbolt-5)). #### Two ways to split a model **Pipeline parallelism** gives each Mac a run of consecutive layers. It lets a model bigger than any single Mac run at all, and long prompts can be processed in parallel chunks, but generation is about as fast as on one Mac. **Tensor parallelism** splits every layer across the Macs. It speeds up generation too, but the Macs exchange data many times per token, so it needs RDMA's low latency. Apple has published measurements of both on Mac Studio clusters ([Apple Machine Learning Research](https://machinelearning.apple.com/research/mlx-distributed-inference)), and states that four Mac Studios run inference up to three times faster than one ([Apple](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/)). #### Building a cluster 1. Choose Macs with Thunderbolt 5 for RDMA: M4 Pro, M4 Max, M3 Ultra and later Pro, Max and Ultra chips. Matching Macs make the split simplest. 2. Connect them with Thunderbolt 5 cables, each pair directly. Two cables between a pair double the link's bandwidth. 3. Update every Mac to macOS 26.2 or later and enable RDMA as Apple's documentation describes. 4. Run the model with MLX distributed (`mlx.launch`) or with [exo](https://github.com/exo-explore/exo), which discovers the Macs and splits the model for you. exo reports tensor-parallel speedups of up to 1.8 times on two devices and 3.2 times on four. Measure your own setup: the gains depend heavily on the model. #### From a cluster to a network A cluster joins a few Macs on one desk with cables. MLX Network joins Macs everywhere over the internet. Because internet latency is far too high to split one model across strangers' Macs, the network places each job whole on a single Mac, or on a pod of Macs that are already cabled together, by memory and bandwidth. If you own a Mac Studio or a stack of Mac minis, [MLX Node](https://mlx.app/download) can lend their spare time to the network and earn MLX Tokens, and an Ultra [dedicated seat](https://mlx.app/nodes) running on an Ultra Mac earns the most. See every Mac on the network on [MLX World](https://world.mlx.app). #### Questions **Can I cluster Mac minis for AI?** Yes. Mac minis with an M4 Pro or M5 Pro chip have Thunderbolt 5 and, with macOS 26.2 or later, can use RDMA; the M4 and M6 models have Thunderbolt 4. MLX distributed and exo both split a model across them. **Does a Mac cluster make a model faster?** It depends on how the model is split. Pipeline parallelism mainly lets bigger models run; tensor parallelism over RDMA also speeds up generation. Apple states up to three times faster on four Mac Studios. **Do I need Thunderbolt 5?** For RDMA, yes. Older Macs can still be joined over Thunderbolt or Ethernet with TCP, but latency is much higher, so pipeline splits work better than tensor splits. ### MLX vs Ollama vs llama.cpp URL: https://mlx.app/learn/mlx-vs-ollama · Updated 2026-10-19 · Run AI on Apple Silicon **The short version** - MLX is Apple's machine learning framework for Apple Silicon. mlx-lm runs language models on it. - llama.cpp is a fast C/C++ inference engine that runs GGUF models on almost any hardware, including Macs through Metal. - Ollama is the easy front end: one command to download and run a model, with a local API. On Apple Silicon it now has an MLX-powered preview. - For the fastest Mac-native results use MLX; for the widest choice of models and platforms, llama.cpp; for the easiest start, Ollama or LM Studio. Three names come up whenever someone runs a language model on a Mac. They overlap more than they compete. Here is what each one is, and when to use which. #### What each one is Sources: [MLX](https://github.com/ml-explore/mlx), [llama.cpp](https://github.com/ggml-org/llama.cpp), [Ollama](https://ollama.com/blog/mlx). | | MLX (mlx-lm) | llama.cpp | Ollama | | --- | --- | --- | --- | | Made by | Apple, open source (MIT) | Open-source community | Ollama, open source | | Runs on | Apple Silicon | Mac, Windows, Linux, many GPUs | Mac, Windows, Linux | | Model format | MLX (safetensors) | GGUF | Its own library; GGUF and, in preview, MLX on Mac | | Best for | Speed and research on Macs | Portability, quantization choice | Ease of use, apps that need an API | | Training and fine-tuning | Yes (LoRA and full) | Limited | No | #### Speed on Apple Silicon All three are limited by the same thing when generating text: memory bandwidth. Differences come from how well each uses the GPU and how its models are quantized. MLX is written for Apple Silicon's unified memory and, on M5 chips, uses the GPU's neural accelerators: Apple measured time to first token 3.3 to 4.1 times faster than on M4 ([Apple](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)). In March 2026 Ollama began moving to MLX on Apple Silicon in a preview with large speedups for supported models ([Ollama](https://ollama.com/blog/mlx)), which says a lot about where Mac performance is. llama.cpp remains very fast and well measured on Macs; its public benchmark thread is the best reference for comparing chips ([llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/4167)). For numbers on MLX specifically, see [MLX Bench](https://mlx.app/bench). #### Which one to use - **You want to chat with a model today:** LM Studio (an app that runs both MLX and GGUF models) or Ollama. - **You're building an app that needs a local API:** Ollama, or `mlx_lm.server` for an OpenAI-compatible endpoint on MLX. - **You want the most speed from a Mac, or to fine-tune:** MLX with mlx-lm. - **You need the same setup on Mac, Linux and Windows, or a specific GGUF quantization:** llama.cpp. Our step-by-step guide to [running an LLM on your Mac](https://mlx.app/learn/run-llm-on-mac) covers all of them. #### MLX beyond one Mac MLX also runs across several Macs, over Thunderbolt 5 with RDMA ([Mac AI clusters](https://mlx.app/learn/mac-ai-cluster)). And it's the engine of MLX Network: [MLX Node](https://mlx.app/download) runs AI jobs on the MLX framework for the network and earns MLX Tokens for verified work. MLX Network is an independent project, not made by Apple; [here's the difference](https://mlx.app/learn/mlx-network-vs-apple-mlx). #### Questions **Is MLX faster than llama.cpp?** On Apple Silicon, MLX is often as fast or faster for generation and especially for prompt processing on M5 chips. Results vary by model and quantization, so compare on your own Mac. **Can Ollama use MLX?** Yes, in preview since Ollama 0.19 on Apple Silicon, for supported models, on Macs with more than 32 GB of memory. **Do MLX and GGUF models differ?** They're different file formats for the same kinds of models. Many popular models are published in both; the mlx-community organization on Hugging Face has thousands of MLX conversions. ### The best apps to run LLMs on a Mac URL: https://mlx.app/learn/best-apps-to-run-llm-on-mac · Updated 2026-10-20 · Run AI on Apple Silicon **The short version** - For Apple Silicon there are two model formats: MLX, made for Apple's chips, and GGUF, from llama.cpp, which runs everywhere. - MLX Harness and mlx-lm run MLX; llama.cpp runs GGUF; LM Studio, Jan and (in preview) Ollama run both. - Fully open source: MLX Harness, Ollama, Jan, llama.cpp and mlx-lm. LM Studio is free, also for work, but its desktop app's source isn't published. - Whatever you choose, memory decides which models fit: [check your Mac](https://mlx.app/mac) first. Six tools cover almost every way to run an open AI model on a Mac. They differ in four things that matter: whether you can read the code, which model formats they run, whether there is an app or only a command line, and what kind of local API your other apps can use. #### At a glance Sources: [MLX Harness](https://github.com/MLXapp/harness), [LM Studio on GitHub](https://github.com/lmstudio-ai), [LM Studio MLX](https://lmstudio.ai/mlx), [Ollama](https://github.com/ollama/ollama), [Ollama MLX preview](https://ollama.com/blog/mlx), [Jan](https://github.com/janhq/jan), [llama.cpp server](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md), [mlx-lm server](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md). | | Open source | Runs on a Mac | App | Local API | | --- | --- | --- | --- | --- | | MLX Harness | Yes (Apache 2.0) | MLX | Mac app | OpenAI, Anthropic and Ollama-style | | LM Studio | CLI, SDKs and MLX engine (MIT); app source not published | MLX and GGUF | Mac app and CLI | OpenAI and Anthropic | | Ollama | Yes (MIT) | GGUF; MLX in preview | Mac app and CLI | Its own, plus OpenAI and Anthropic subsets | | Jan | Yes (Apache 2.0) | GGUF and MLX | Mac app | OpenAI | | llama.cpp | Yes (MIT) | GGUF | Command line and a web UI | OpenAI and Anthropic | | mlx-lm | Yes (MIT) | MLX | Command line | OpenAI-style, basic | #### Each one in a few lines ##### MLX Harness A free, open-source Mac app for chat, coding agents with built-in tools, image generation and a local API that speaks the OpenAI, Anthropic and Ollama formats. It runs models with Apple's MLX and downloads them from Hugging Face; it doesn't run GGUF files. It needs macOS 14.5 or later on Apple Silicon. It's ours, built on vMLX by JANGQ AI, and it carries [MLX Node](https://mlx.app/download), which can put the Mac to work for MLX Network while you're away ([source](https://github.com/MLXapp/harness)). Guide: [how to use MLX Harness](https://mlx.app/learn/how-to-use-mlx-harness). ##### LM Studio A polished app that runs both MLX and GGUF models side by side, with a model search built on Hugging Face ([LM Studio](https://lmstudio.ai/mlx)). It is free to use at home and at work since July 2025 ([LM Studio](https://lmstudio.ai/blog/free-for-work)). Its command line tool, SDKs and MLX engine are MIT-licensed; the desktop app's source isn't published ([GitHub](https://github.com/lmstudio-ai)). It needs macOS 14 or later on Apple Silicon, and recommends 16 GB of memory ([requirements](https://lmstudio.ai/docs/app/system-requirements)). Its local server speaks the OpenAI format and, since version 0.4.1, the Anthropic format ([changelog](https://lmstudio.ai/changelog/lmstudio-v0.4.1)). ##### Ollama The one-command way to run a model, with its own model library and, since July 2025, a chat app on the Mac ([Ollama](https://ollama.com/blog/new-app)). It's MIT-licensed and runs GGUF models through llama.cpp ([GitHub](https://github.com/ollama/ollama)). Since March 2026 it has an MLX engine in preview on Apple Silicon, for Macs with more than 32 GB of memory and a growing list of models ([Ollama](https://ollama.com/blog/mlx)). It needs macOS 14 or later ([docs](https://docs.ollama.com/macos)), and its API on port 11434 is what many other apps expect. ##### Jan An open-source (Apache 2.0) chat app for Mac, Windows and Linux, with a bundled llama.cpp engine, native MLX support since February 2026, MCP tools, and an OpenAI-compatible server ([GitHub](https://github.com/janhq/jan), [changelog](https://www.jan.ai/changelog/2026-02-11-jan-mlx-ui-refresh)). It can also connect to cloud models, if you want one app for both. ##### llama.cpp The engine underneath much of local AI: MIT-licensed C/C++, with Apple Silicon as a "first-class citizen" through Metal ([GitHub](https://github.com/ggml-org/llama.cpp)). Install it with `brew install llama.cpp`; `llama-server` gives you a web chat and an OpenAI- and Anthropic-compatible API ([server docs](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md)). It runs GGUF files, the most widely published format. ##### mlx-lm Apple's own MLX tools for language models, MIT-licensed: generate, chat, convert and quantize models, and fine-tune them with LoRA, from the command line ([GitHub](https://github.com/ml-explore/mlx-lm)). `mlx_lm.server` gives a basic OpenAI-style endpoint that its authors don't recommend for production ([server docs](https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md)). Thousands of converted models are in the mlx-community organization on Hugging Face. > Note: GPT4All (MIT) still installs on Macs, but it has had no new release since February 2025 ([releases](https://github.com/nomic-ai/gpt4all/releases)). #### Which one to pick - **You want an app, open source, made for Apple Silicon:** [MLX Harness](https://harness.mlx.app/). - **You want MLX and GGUF models in one polished app, and don't mind closed source:** LM Studio. - **Another app or coding agent expects the Ollama API:** Ollama (MLX Harness also serves Ollama-style endpoints). - **You want one chat app across Mac, Windows and Linux, local and cloud:** Jan. - **You script, build servers or need a particular GGUF quantization:** llama.cpp. - **You fine-tune, convert models or want Apple's reference implementation:** mlx-lm. Many people use two: an app for everyday chat and a command line tool for scripts. They can share the same Mac, but each keeps its own copy of the model files. #### First, check what your Mac can hold No app changes the arithmetic: a model and its context have to fit in the share of unified memory macOS lets the GPU use, about 75% by default, and memory bandwidth sets how fast it writes. MLX and GGUF files of the same model at the same precision are about the same size. See every model that fits [your Mac](https://mlx.app/mac), the [best models for each memory size](https://mlx.app/mac/16gb), or start from [a model](https://mlx.app/run) to see which Mac it needs. #### When the Mac is idle A Mac that runs local AI well is also good at running it for others. [MLX Node](https://mlx.app/download) runs AI jobs for MLX Network when you're not using your Mac, within the limits you set, and earns MLX Tokens for verified work. MLX Harness includes it. MLX Tokens can lose value; read the [risks](https://mlx.app/risks). #### Questions **Is LM Studio free for work?** Yes. Since July 2025 LM Studio is free to use both at home and at work; paid plans are for teams and enterprises ([LM Studio](https://lmstudio.ai/blog/free-for-work)). **Does Ollama use MLX on a Mac?** In preview since March 2026, on Apple Silicon Macs with more than 32 GB of memory, for supported models. Otherwise it runs GGUF models through llama.cpp ([Ollama](https://ollama.com/blog/mlx)). **Which app runs both MLX and GGUF models?** LM Studio and Jan run both; Ollama runs GGUF and, in preview, MLX. MLX Harness and mlx-lm run MLX only; llama.cpp runs GGUF only. **Which local LLM apps are open source?** MLX Harness (Apache 2.0), Ollama (MIT), Jan (Apache 2.0), llama.cpp (MIT) and mlx-lm (MIT). LM Studio publishes its command line tool, SDKs and MLX engine under MIT, but not the desktop app. ### Can my Mac run Perplexity Computer locally? URL: https://mlx.app/learn/run-perplexity-computer-locally-on-mac · Updated 2026-10-20 · Run AI on Apple Silicon **The short version** - Perplexity Computer can now run the model, the agent and its tasks on a supported Apple silicon Mac, for Pro, Max and Enterprise subscribers ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac)). - Perplexity recommends 36 GB of unified memory or more. Macs with a Pro, Max or Ultra chip can have that; Macs with a base M chip top out at 32 GB. - The local models are Qwen 3.6 and PPLX Pro, a post-trained version of Qwen. Perplexity hasn't said how big PPLX Pro is. - Check your Mac's memory in the Apple menu under About This Mac, then see [what a 36 GB Mac runs](https://mlx.app/mac/36gb). On 20 October 2026 Perplexity made its Computer agent run entirely on a Mac: the AI model, the agent and the work it does. Whether your Mac is a good fit comes down mostly to one number, its unified memory. #### What Perplexity announced Perplexity Computer is Perplexity's agent that carries out tasks across your apps and files. On 20 October 2026 Perplexity said it can now run fully local on supported Apple silicon Macs, "making it possible to run the AI model, harness, and task execution entirely on device". Planning, reasoning and file processing then stay on the Mac, and code and tools run in isolated sandboxes. It's for Perplexity Pro, Max and Enterprise subscribers, and it can still hand a task to the cloud when you allow it ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac)). Two local models come with it: Qwen 3.6 and PPLX Pro, which Perplexity describes as a post-trained version of Qwen. Its own example of a supported Mac is the new Mac mini with M6 and M5 Pro ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac)). [9to5Mac](https://9to5mac.com/2026/10/20/perplexitys-computer-can-now-run-entirely-on-your-mac-even-offline/) and [AppleInsider](https://appleinsider.com/articles/26/10/20/perplexity-computer-can-run-completely-local-on-mac-with-a-big-caveat) covered the launch. #### Why 36 GB, and which Macs have it For running Computer locally, Perplexity writes, "36 GB or more is recommended". That covers the model, its context and the agent around it; Perplexity doesn't break the figure down, and it's a recommendation, not a hard limit ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac)). Memory options from Apple's Tech Specs, for example the [Mac mini (M6 or M5 Pro)](https://support.apple.com/en-us/128108) and [Mac Studio (M5 Max or M5 Ultra)](https://support.apple.com/en-us/128107). Every chip with its sources: [Macs by chip](https://mlx.app/mac). | Chip | Memory you can order | 36 GB or more? | | --- | --- | --- | | M6 | 16, 24 or 32 GB | No | | M5 | 16, 24 or 32 GB | No | | M5 Pro | 24, 48 or 64 GB | With 48 or 64 GB | | M5 Max | 36, 48, 64 or 128 GB | Yes | | M5 Ultra | 96, 256 or 512 GB | Yes | | M4 | 16, 24 or 32 GB | No | | M4 Pro | 24, 48 or 64 GB | With 48 or 64 GB | | M4 Max | 36, 48, 64 or 128 GB | Yes | | M3 Pro | 18 or 36 GB | With 36 GB | | M3 Max | 36, 48, 64, 96 or 128 GB | Yes | | M3 Ultra | 96, 256 or 512 GB | Yes | Older Macs count too: the M1 Max with 64 GB, the M2 Max with 64 or 96 GB, and every M1, M2 and M3 Ultra. Every Mac with a base M chip, and the M1 Pro and M2 Pro, top out at 32 GB or less. #### What the GPU actually gets A Mac's memory is shared by the CPU and the GPU, and macOS lets the GPU use only part of it by default. Our [Compare](https://mlx.app/compare) page counts about 75%: about 27 GB for a model and its context on a 36 GB Mac, about 36 GB on a 48 GB Mac and about 24 GB on a 32 GB Mac. mlx-lm's README shows how to raise the limit on macOS 15 or later ([mlx-lm](https://github.com/ml-explore/mlx-lm)), but macOS and your other apps still need room. Perplexity hasn't said which size of Qwen 3.6 the local mode uses. The one it benchmarked on a Mac in September is Qwen3.6-35B-A3B: 35 billion parameters with about 3 billion active per token, a 19.4 GB 4-bit checkpoint, and 186.6 tokens a second at a 4K context on an M5 Max MacBook Pro with 128 GB, using Perplexity's own Lily engine ([Perplexity](https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon)). Our own arithmetic for that model at 4-bit, 19.7 GB with MLX's scales, agrees on the size: it fits a 32 GB Mac with a little room left, fits a 36 GB Mac comfortably, and fits a 24 GB Mac only at 3-bit or with the GPU limit raised. See [Qwen3.6 35B A3B on every Mac](https://mlx.app/run/qwen3.6-35b-a3b). #### If your Mac has 24 or 32 GB Since 36 GB is a recommendation, you may still be able to try it, but expect less room for long tasks next to your other apps. Local AI in general still works well at these sizes: see [what a 32 GB Mac runs](https://mlx.app/mac/32gb), [what a 24 GB Mac runs](https://mlx.app/mac/24gb) and the [Mac mini with M6](https://mlx.app/mac/mac-mini-m6). If you're buying for it, a Pro chip with 48 GB gets you past the recommendation, for example the [Mac mini with M5 Pro](https://mlx.app/mac/mac-mini-m5-pro) or an [M4 Pro with 48 GB](https://mlx.app/mac/m4-pro-48gb). A Max chip starts at 36 GB and reads memory faster, which makes models write faster. Compare [36 GB](https://mlx.app/mac/36gb) and [48 GB](https://mlx.app/mac/48gb) Macs, or read [the best Mac for local LLMs](https://mlx.app/learn/best-mac-for-local-llm). #### Other ways to use AI on your own Mac Perplexity Computer is one way to run an agent locally. [MLX Harness](https://harness.mlx.app/) is our free, open-source Mac app for chat, coding agents with built-in tools and a local API, running open models with Apple's MLX ([how to use it](https://mlx.app/learn/how-to-use-mlx-harness)). [The best apps to run LLMs on a Mac](https://mlx.app/learn/best-apps-to-run-llm-on-mac) compares it with LM Studio, Ollama and others. A Mac with this much memory is also idle most of the day. [MLX Node](https://mlx.app/download) runs AI jobs for MLX Network when you're not using it, within the limits you set, and earns MLX Tokens for verified work. MLX Tokens can lose value; read the [risks](https://mlx.app/risks). MLX Network is independent and not affiliated with Perplexity or Apple. #### Questions **Does Perplexity Computer work offline on a Mac?** Perplexity says the model, harness and task execution can run entirely on the device, so personal data stays on the Mac, and 9to5Mac reports that tasks can run without an internet connection ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac), [9to5Mac](https://9to5mac.com/2026/10/20/perplexitys-computer-can-now-run-entirely-on-your-mac-even-offline/)). It can still hand a task to the cloud when you allow it. **Is 32 GB enough for Perplexity Computer?** It's below Perplexity's recommendation of 36 GB or more. It's a recommendation rather than a hard minimum, so a 32 GB Mac may still run it, with less room for the agent's work next to your other apps. **What is PPLX Pro?** One of the two models Perplexity Computer can run locally, which Perplexity describes as a post-trained version of Qwen. Perplexity hasn't published its size. **Which Mac mini can run Perplexity Computer locally?** The Mac mini with M5 Pro, ordered with 48 or 64 GB, meets the 36 GB recommendation. The Mac mini with M6 tops out at 32 GB ([Apple](https://support.apple.com/en-us/128108)). **Do I need a Perplexity subscription?** Yes. The local mode is for Perplexity Pro, Max and Enterprise subscribers ([Perplexity](https://www.perplexity.ai/hub/blog/perplexity-computer-now-runs-fully-local-on-mac)). ### The macOS fm command: Apple's local LLM in Terminal URL: https://mlx.app/learn/macos-fm-command-free-local-llm · Updated 2026-10-21 · Run AI on Apple Silicon **The short version** - fm comes pre-installed with macOS 27. Type `fm` in Terminal to list its commands, such as `respond`, `chat` and `schema` ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). - By default it uses the on-device model that comes with macOS. `--model pcc` sends a prompt to Apple's much bigger Private Cloud Compute model instead, which has usage limits. - The on-device model needs a Mac that supports Apple Intelligence (M1 or later). Apple Intelligence takes up to 14 GB of storage on Macs with M3 or later and 12 GB of memory or more, and up to 8 GB on the others ([Apple](https://support.apple.com/en-us/121115)). - Its context window is small: 4,096 tokens in Apple's documentation, 8,192 in a WWDC26 code sample. For long documents or harder tasks, an open model on your Mac's GPU goes further: see [what your Mac can run](https://mlx.app/mac). macOS 27 comes with a command line tool called fm that puts Apple's own language model in Terminal: one-off prompts, a chat, structured JSON for scripts. Here's what Apple says it does, what your Mac needs, where the small on-device model stops, and what an open model on the same Mac can do instead. #### What fm is fm is Apple's command line tool for Apple Foundation Models, the models behind Apple Intelligence. "Starting from macOS 27, this command line tool comes pre-installed on your Mac," Apple says in its WWDC26 session on it; you open Terminal and type `fm` ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). Apple pitches it for scripting and automation "without writing a line of app code" ([Apple](https://developer.apple.com/foundation-models-framework/)). It can talk to two models. By default it uses "the on-device model that comes with macOS, and that's always available". It can also use the Apple Foundation Model on Private Cloud Compute, which Apple calls "a much bigger model than the on-device model" that "will perform better on complex problems" and "has usage limits" ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). The on-device model runs on your Mac and works without an internet connection ([Apple](https://developer.apple.com/foundation-models-framework/)); a prompt sent to Private Cloud Compute goes to Apple's servers. fm is Apple's software and has nothing to do with MLX Network. MLX Network is independent and not affiliated with Apple. #### The commands These are the commands and options Apple shows in its WWDC26 sessions ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/), [Apple](https://developer.apple.com/videos/play/wwdc2026/241/)): Apple says there are more commands than these; `fm` alone lists them on your Mac. | Command or option | What it does | | --- | --- | | `fm` | Lists the commands that are available. | | `fm respond "…"` | Prompts the model and prints its response, for scripts and one-off questions. | | `fm chat` | An interactive conversation in Terminal. Inside it, `/model` switches to the Private Cloud Compute model and `/save` saves the conversation to resume later. | | `fm schema object` | Creates a schema, so `fm respond` returns structured JSON. | | `--model pcc` | Sends the prompt to the Private Cloud Compute model instead of the on-device one. | | `--image` | Adds an image to the prompt. | | `--schema` | Makes the answer follow a schema made with `fm schema object`. | | `--instructions` | Gives the model instructions separate from the prompt. | | `--help` | Shows every option. | Apple's own examples, from the session: ``` fm respond "Provide a basic regex in Swift to parse an email address" fm respond "Provide a comprehensive regex in Swift to parse an email address" --model pcc fm schema object --name AppsIdentified --string app_names --array > schema.json fm respond "What apps are the user actively using in this screenshot?" \ --image Screenshot.png --model pcc --schema schema.json # {"app_names": ["Messages", "Mail", "Calendar"]} fm respond --help ``` > Note: Posts about macOS 27 also describe commands such as `fm serve`, `fm available` and `fm license`. Apple's WWDC26 sessions and the Foundation Models documentation we read don't describe them, so this guide doesn't either. Run `fm` on your Mac to see what your version has. #### What you need, and how much it downloads - **macOS 27.** fm comes with it ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). - **A Mac that supports Apple Intelligence.** Apple Foundation Models need one ([Apple](https://developer.apple.com/documentation/foundationmodels)); on macOS 27 that's any Mac with M1 or later ([Apple](https://support.apple.com/en-us/121115)). - **A supported language.** Apple Intelligence needs the device language and Siri language set to the same supported language, and it doesn't currently work on devices bought in mainland China ([Apple](https://support.apple.com/en-us/121115)). - **Free disk space** for Apple Intelligence, below. From Apple's [Apple Intelligence requirements](https://support.apple.com/en-us/121115). Apple gives these figures for Apple Intelligence as a whole, not for the model fm uses alone. | Mac | Apple Intelligence storage on macOS 27 | | --- | --- | | M3 or later with at least 12 GB of unified memory | Up to 14 GB | | Other Macs with Apple Intelligence (M1, M2, or less than 12 GB of memory) | Up to 8 GB | The model can be unavailable for a while after you turn Apple Intelligence on: Apple's framework reports it as not ready "because it's downloading or because of other system reasons" ([Apple](https://developer.apple.com/documentation/foundationmodels/systemlanguagemodel)). For the Python SDK, which uses the same on-device model, Apple lists macOS 26.0 or later, Xcode 26.0 or later with its agreement accepted, and Python 3.10 or later ([Apple](https://github.com/apple/python-apple-fm-sdk)), on an Apple silicon Mac ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). #### What the on-device model is good at, and where it stops Apple is plain about it: "On-device models excel at a diverse range of text generation tasks, like summarization, entity extraction, text and image understanding, refinement, dialog for games, generating creative content, and more. When you need more reasoning capabilities and context size, use Private Cloud Compute or any server model provider" ([Apple](https://developer.apple.com/documentation/foundationmodels)). - **A small context window.** Apple's documentation gives the on-device model "a context window of 4096 tokens per session" ([Apple](https://developer.apple.com/documentation/foundationmodels/managing-the-context-window)) and lists 4K against 32K for Private Cloud Compute ([Apple](https://developer.apple.com/documentation/foundationmodels/adding-server-side-intelligence-with-private-cloud-compute)). In its WWDC26 session on macOS 27, Apple's code sample prints a `contextSize` of 8,192 ([Apple](https://developer.apple.com/videos/play/wwdc2026/241/)). The prompt, the instructions and the answer all share that window. - **What that holds.** In English a token is typically three to four characters ([Apple](https://developer.apple.com/documentation/foundationmodels/managing-the-context-window)), so 8,192 tokens is roughly 25,000 to 33,000 characters in and out together: a few pages, not a book or a codebase. - **Short, simple prompts.** "An on-device model has fewer parameters and a small context window," so Apple asks for prompts that are "short and succinct", and says on-device models "have limited reasoning abilities": break a hard task into steps ([Apple](https://developer.apple.com/documentation/foundationmodels/prompting-an-on-device-foundation-model)). - **It changes with macOS.** Apple updates the model in OS updates; macOS 27 has a new version ([Apple](https://developer.apple.com/documentation/foundationmodels/systemlanguagemodel)). A script's results can shift after an update. Apple doesn't give the on-device model's size on these pages, so we don't either. #### Using fm from scripts and code `fm respond` prints the answer, so it drops into any shell script. With `--schema` the answer is JSON you can pass to `jq`. Apple's session sorts a project folder this way: it asks the model which files are final and which are drafts, then copies and moves them. Shortened from Apple's sample ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)): ``` fm schema object --name "TriagedFileList" \ --string 'final_files' --array \ --string 'draft_files' --array > /tmp/schema.json output=$(fm respond \ --instructions "Return the latest files, then separately all draft files." \ "This is the list of all files:\n\n${files_list}" \ --schema /tmp/schema.json ) echo "${output}" | jq -r '.draft_files[]' ``` For Python there's Apple's Foundation Models SDK, `pip install apple_fm_sdk`, with the same on-device model, guided generation and tool calling ([Apple](https://github.com/apple/python-apple-fm-sdk)). Apple's example: ``` import apple_fm_sdk as fm model = fm.SystemLanguageModel() is_available, reason = model.is_available() if is_available: session = fm.LanguageModelSession(model=model) response = await session.respond(prompt="Hello!") print(response) ``` Many apps and coding agents expect an OpenAI- or Anthropic-style API on a local port instead. If that's what you need, [MLX Harness](https://mlx.app/learn/how-to-use-mlx-harness) serves every model you start through one gateway at http://127.0.0.1:8080 in OpenAI, Anthropic and Ollama formats, with open models of your choice. More options: [the best apps to run LLMs on a Mac](https://mlx.app/learn/best-apps-to-run-llm-on-mac). #### When to switch to an open model Stay with fm for short jobs: summarising a note, naming a file, pulling fields out of text, a quick question. It's already on the Mac and needs no setup. Reach for an open model when the input won't fit in a few thousand tokens, when the task needs more reasoning or code, when you want to pick the model, or when a tool needs an API. An open model runs on your Mac's GPU, and what fits is set by unified memory. Our [Compare](https://mlx.app/compare) arithmetic gives the GPU about 75% of it by default and leaves room for the context: Our arithmetic on the models' official cards: MLX stores (bits + 0.5) per weight, plus 8% or 1.5 GB for context. Everything each size runs, with speed ranges: [8 GB](https://mlx.app/mac/8gb), [16 GB](https://mlx.app/mac/16gb), [24 GB](https://mlx.app/mac/24gb), [32 GB](https://mlx.app/mac/32gb). | Mac memory | For the GPU | Runs well, for example | Memory it needs | | --- | --- | --- | --- | | 8 GB | about 6 GB | [Llama 3.1 8B](https://mlx.app/run/llama-3.1-8b) at 4-bit | 4.5 GB, 6.0 GB with context | | 16 GB | about 12 GB | [Llama 3.1 8B](https://mlx.app/run/llama-3.1-8b) at 8-bit, [Qwen3.5 9B](https://mlx.app/run/qwen3.5-9b) at 6-bit | 8.5 GB and 7.3 GB, 10.0 and 8.8 GB with context | | 24 GB | about 18 GB | [gpt-oss-20b](https://mlx.app/run/gpt-oss-20b), native 4.25-bit | 13.8 GB, 15.3 GB with context | | 32 GB | about 24 GB | [Qwen3.6 35B A3B](https://mlx.app/run/qwen3.6-35b-a3b) at 4-bit | 19.7 GB, 21.3 GB with context | These models list much longer context windows on their model cards, 128K tokens for [Llama 3.1 8B](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) and 262,144 for [Qwen3.6 35B A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B), though a long context needs memory beyond the room counted above. Not sure what your Mac has? Apple menu, About This Mac, then see [Macs by chip](https://mlx.app/mac). [MLX Harness](https://harness.mlx.app/) is one way to run them: our free, open-source Mac app, built on vMLX by JANGQ AI, finds MLX models on Hugging Face, chats, runs coding agents with built-in tools and serves the local API above ([how to use it](https://mlx.app/learn/how-to-use-mlx-harness)). When you're not using the Mac, [MLX Node](https://mlx.app/download) can run AI jobs for MLX Network within the limits you set and earns MLX Tokens for verified work. MLX Tokens can lose value; read the [risks](https://mlx.app/risks). #### Questions **Is the macOS fm command free?** It comes pre-installed with macOS 27 and by default uses the on-device model that comes with macOS. The bigger Private Cloud Compute model has usage limits ([Apple](https://developer.apple.com/videos/play/wwdc2026/334/)). **Does fm work offline?** With the on-device model, yes: Apple says it works without an internet connection ([Apple](https://developer.apple.com/foundation-models-framework/)). Apple Intelligence has to be downloaded first, and `--model pcc` or `/model` in `fm chat` use Apple's servers. **How long a prompt can fm take?** Prompt, instructions and answer share one context window: 4,096 tokens in Apple's documentation and 8,192 in Apple's WWDC26 sample for macOS 27 ([Apple](https://developer.apple.com/videos/play/wwdc2026/241/)). In English that's roughly three to four characters a token. **Does fm serve an OpenAI-compatible API?** Apple's WWDC26 sessions and the Foundation Models documentation we read don't describe a server mode, so we can't confirm how one works. For a documented local API, [MLX Harness](https://mlx.app/learn/how-to-use-mlx-harness) serves open models at http://127.0.0.1:8080 in OpenAI, Anthropic and Ollama formats. **Which Macs can use fm?** Macs on macOS 27 that support Apple Intelligence: any Mac with M1 or later, set to a supported language, with up to 14 GB (M3 or later with 12 GB or more) or up to 8 GB of free storage ([Apple](https://support.apple.com/en-us/121115)). ### What is DePIN? URL: https://mlx.app/learn/what-is-depin · Updated 2026-10-19 · MLX Network **The short version** - In a DePIN, many independent owners supply hardware such as storage, wireless coverage, sensors or compute, and are paid in a token. - The hard problems are proving the work was really done and making the token worth something beyond rewards. - Decentralized AI compute is a natural fit, because AI needs more compute than data centers can supply. - MLX Network is a DePIN for Apple Silicon Macs: Macs run AI work and earn MLX Tokens. DePIN stands for decentralized physical infrastructure networks: networks where ordinary people supply the hardware, and a token pays them for it. Here's how they work, what can go wrong, and what decentralized AI compute looks like. #### DePIN in one paragraph Traditional infrastructure is built by one company that pays for all of it up front. A DePIN does it the other way round: it publishes the rules, and anyone who supplies the right hardware joins and is paid in the network's token. The token is the incentive to build before there's revenue; customers paying for the service are what should sustain it later. The idea goes back to volunteer computing. Folding@home pooled home computers into one of the most powerful computing systems in the world in 2020 ([Wikipedia](https://en.wikipedia.org/wiki/Folding@home)). Bitcoin added the payment: a token that rewards whoever supplies the work ([Nakamoto](https://bitcoin.org/bitcoin.pdf)). #### Kinds of DePIN - **Storage:** people rent out disk space. - **Wireless:** people run hotspots and are paid for coverage. - **Sensors and maps:** dashcams, weather stations and location data. - **Bandwidth:** people share their internet connection. - **Compute:** people supply processors and graphics for rendering, training or AI inference. #### What makes a DePIN work, or fail **Verification.** The network has to know the work was real. Storage can be checked with proofs; AI is harder, because a cheap wrong answer looks much like an expensive right one. Research ranges from cryptographic proofs, which are still too slow for large models ([zkLLM](https://arxiv.org/abs/2404.16109)), to comparing answers across machines and spot checks ([TOPLOC](https://arxiv.org/abs/2501.16007)). **Real demand.** Token rewards can attract supply for a while, but a network lasts only if someone pays for what it provides. **Token risk.** Rewards are paid in a token whose price moves. Early participants are often paid most, which is also why the risks deserve a careful read. #### Decentralized AI compute, and Macs AI needs more compute than data centers can supply: the largest tech companies planned more than $700 billion in capital spending for 2026 ([Sherwood News](https://sherwood.news/tech/alphabet-amazon-microsoft-meta-plan-more-than-700-billion-on-capex-this-year/)). Most compute DePINs pool Nvidia graphics cards. MLX Network pools Apple Silicon Macs, whose large unified memory suits running models whole. How it works: [MLX Node](https://mlx.app/download) runs AI jobs on Apple's MLX framework; answers are checked against other Macs and known test jobs; every hour the network pays MLX Tokens from a fixed compute pool, locked for 30 days on a free seat and 3 to 20 with a dedicated seat. [Dedicated seats](https://mlx.app/nodes) work like a stake: each reserves a Mac's place on the network and multiplies its share of the rewards; a seat pays nothing up front. Every block and transaction is public on the [explorer](https://explorer.mlx.app), every Mac on [MLX World](https://world.mlx.app), and the rules are in the [whitepaper](https://mlx.app/whitepaper). #### Questions **What does DePIN stand for?** Decentralized physical infrastructure network: a network built from hardware supplied by many independent owners, who are paid in a token. **Is DePIN a good investment?** We can't tell you that, and nobody can promise it. DePIN tokens are volatile and can lose all their value. Judge a network by whether real customers pay for what it provides, and read its risk disclosures. **Is MLX a DePIN?** Yes. MLX Network is a decentralized compute network for Apple Silicon: Mac owners supply the compute with MLX Node and earn MLX Tokens for verified AI work. ### MLX Network and Apple's MLX URL: https://mlx.app/learn/mlx-network-vs-apple-mlx · Updated 2026-10-19 · MLX Network **The short version** - MLX (by Apple) is an open-source array framework for machine learning on Apple Silicon, on GitHub as ml-explore/mlx. - MLX Network is an independent community project: Macs run AI work with the MLX framework and earn the MLX Token. - MLX Network is not affiliated with Apple Inc. Apple's framework is free and doesn't involve any token. - If you came for the framework, Apple's documentation and the mlx-community models on Hugging Face are the place to start. Two things share the name MLX. One is Apple's open-source framework for machine learning on Apple Silicon. The other is MLX Network, the independent compute network at mlx.app that runs on it. Here's the difference. #### Apple's MLX framework MLX is an array framework for machine learning on Apple Silicon, released by Apple's machine learning research team under the MIT license ([GitHub](https://github.com/ml-explore/mlx)). It feels like NumPy and PyTorch, uses the Mac's unified memory without copying between processor and graphics, and powers **mlx-lm** for language models, MLX Swift for apps and distributed inference across several Macs. It's free and open source. You don't need an account, a wallet or a token to use it. Start with [running an LLM on your Mac](https://mlx.app/learn/run-llm-on-mac). #### MLX Network MLX Network (this site) is an independent community building the largest Apple Silicon compute network. **MLX Node** runs AI work on the MLX framework within the limits its owner sets, and earns the **MLX Token** for verified results. Dedicated seats multiply what a Mac earns, the ledger is public on the [explorer](https://explorer.mlx.app), and every Mac on the network is on [MLX World](https://world.mlx.app). We chose the name because the network is built on, and for, the MLX framework. MLX Network gives back to that community through [MLX Builders](https://mlx.app/builders): grants and free Builder seats for people who build with MLX, and [MLX Bench](https://mlx.app/bench), an open benchmark of MLX on every Apple Silicon chip. #### Side by side | | MLX (Apple) | MLX Network (mlx.app) | | --- | --- | --- | | What it is | Machine learning framework | Compute network and token | | Made by | Apple's machine learning research team | Independent community, not affiliated with Apple | | Cost | Free, open source | Free app; optional dedicated seats | | Token | None | MLX Token, 10,000,000,000 fixed supply | | Where | github.com/ml-explore/mlx | mlx.app | #### Questions **Is MLX Network made by Apple?** No. MLX Network is an independent community project and is not affiliated with Apple Inc. It runs on Apple's open-source MLX framework. **Is the MLX token related to Apple's MLX?** No. Apple's MLX framework has no token. The MLX Token belongs to MLX Network and pays Macs for AI work. **Where do I download Apple's MLX?** Install it with pip install mlx, or mlx-lm for language models. The source and documentation are on GitHub at ml-explore/mlx. ### Is MLX safe on my Mac? URL: https://mlx.app/learn/is-mlx-safe · Updated 2026-10-19 · MLX Network **The short version** - MLX Node is signed with an Apple Developer ID, notarized by Apple and runs in the macOS App Sandbox. - It can reach the network and its own files. It can't read your documents, photos or other apps. - It works only within your limits and can pause on battery or heat. New installs start in manual mode: it works until you press Stop. - The main risks are financial, not technical: MLX Tokens can lose value. Read the risk disclosure. Letting an app run work on your Mac is a fair thing to be careful about. Here's exactly what MLX Node does, what it can and can't touch, what it sends, and the risks that remain. #### Signed, notarized and sandboxed MLX Node is signed with an Apple Developer ID and notarized: Apple scanned it for malicious software before macOS lets it open ([Apple](https://developer.apple.com/documentation/security/notarizing-macos-software-before-distribution)). Only notarized builds are published, and the app installs an update only after checking it carries the same Developer ID signature. It runs in the macOS App Sandbox. The only permissions it asks for are network access and its Downloads folder, used for updates and for files you choose to save from your wallet. It can't open your documents, photos, mail or other apps' data. A small built-in updater runs outside the sandbox; its only job is to install MLX Node updates signed by the same developer. #### What it runs The app runs AI jobs: prompts for a small open language model, run with Apple's MLX framework, and sends back the answers. The network compares them with other Macs' answers to the same jobs and pays MLX Tokens for verified work. You decide when it works. New installs start in manual mode: once your Mac is linked and plugged in, MLX Node works within your limits until you press Stop. You can instead let it work only after the Mac has been idle, cap how hard it works, and pause it automatically on battery or when the Mac gets warm. MLX.app for iPhone and iPad can pause and resume your Macs remotely. #### What it sends A linked Mac sends a signed status about once a minute: app and macOS version, chip, memory, whether it's working or paused and why, temperature level, whether it's on power, your limits and its measured speed. Each Mac has its own device key in the Secure Enclave, which never leaves the Mac ([Apple](https://developer.apple.com/documentation/security/protecting-keys-with-the-secure-enclave)). Each Mac appears on the public map at [MLX World](https://world.mlx.app) at its approximate city, rounded to about 10 km, without your wallet, the Mac's name or its IP address. The [Privacy Policy](https://mlx.app/privacy) lists everything kept and for how long. #### The risks that remain - **Token value.** MLX Tokens are new. Their price can fall, and they could become worthless. Mined MLX is locked for 30 days (3 to 20 with a dedicated seat). - **An early network.** The ledger is run by the MLX Network operator today, and audits are still to come. The [whitepaper](https://mlx.app/whitepaper) and [risk disclosure](https://mlx.app/risks) explain what that means. - **Wear and power.** Sustained work uses electricity and, on a laptop, battery cycles. Use the limits, or run it on power. Questions? Write to [support@mlx.app](mailto:support@mlx.app) or see [Help](https://mlx.app/support). #### Questions **Is MLX Node a virus or malware?** No. It's signed with an Apple Developer ID, notarized by Apple, and runs in the App Sandbox, which stops it from reading your files or other apps. **Can MLX Node see my files?** No. The sandbox limits it to the network, its own container and its Downloads folder access, used for updates and files you save from your wallet. **How do I stop or uninstall MLX Node?** Press Stop in the menu bar panel, or quit the app. To uninstall, quit it and drag the app from Applications to the Trash. **Is MLX a scam?** Judge for yourself: the app is free, every reward and transaction is public on the explorer, every Mac is on the live map and the rules are in the whitepaper. The token carries real risk, which the risk disclosure sets out plainly. ### How the MLX wallet works URL: https://mlx.app/learn/how-the-mlx-wallet-works · Updated 2026-10-19 · MLX Network **The short version** - Your wallet's secret is 32 random bytes made on your device and stored in your passkey. The signing key is derived from it; no server ever holds it. - Every transaction is signed by your key and recorded on the MLX ledger, which anyone can read on the explorer. - USDC is the network dollar: deposits in any supported coin arrive as USDC, and withdrawals pay USDC on Solana. - The operator runs the ledger and holds the reserves behind USDC on MLX, but can't sign for you. The MLX wallet is a web wallet with no password and no server-side key. Here's what happens when you create one, what each transaction is, where your balances live, and what the MLX Network operator can and can't do with them. #### From passkey to signing key 1. When you create a wallet, your browser makes 32 random bytes: the wallet's seed. 2. The seed is saved as the user handle of a new passkey for mlx.app ([WebAuthn](https://www.w3.org/TR/webauthn-3/)), which your passkey manager keeps and syncs. The browser releases it to the wallet only after Face ID or Touch ID. 3. The wallet turns the seed into an Ed25519 signing key with HKDF-SHA-256 ([RFC 5869](https://www.rfc-editor.org/rfc/rfc5869), [RFC 8032](https://www.rfc-editor.org/rfc/rfc8032)). Its public key is your account's key on the ledger. 4. The same seed, written in the English BIP-39 word list, is your 24-word recovery phrase ([BIP-39](https://github.com/bitcoin/bips/blob/master/bip-0039.mediawiki)). Restoring from it gives the same key. MLX.app for iPhone and iPad derives the key the same way from the same passkey or phrase. While the web wallet is unlocked, the seed stays in that tab's session memory until the wallet locks or the tab closes. > Note: WebAuthn doesn't treat a passkey's user handle as a secret, so the wallet's protection rests on your passkey manager and the Face ID or Touch ID check, not on a key sealed in hardware. The [whitepaper](https://mlx.app/whitepaper#ledger-accounts) says so plainly. Keep your passkey account and recovery phrase safe. #### Names, transactions and the ledger Your account has a readable name like yourname.mlx and the public key from your seed. Every change you make is a transaction: its SHA-256 hash is signed with your key, and the ledger rejects any transaction your key didn't sign. The MLX Network sequencer orders transactions and seals them into hash-linked blocks, each signed by the sequencer, about once a minute. Every account, block and transaction is public on [explorer.mlx.app](https://explorer.mlx.app): names, balances, dedicated seats, transfers, trades and withdrawal addresses. #### Balances, deposits, buying and withdrawing MLX amounts are counted to six decimal places, and so are USDC amounts. **USDC is the network dollar**: it's what you pay with for dedicated seats and spot MLX, and what you withdraw. From the wallet and the ledger's rules. Blockchain payments can't be reversed: check the coin, network and address. | Action | How it works | | --- | --- | | **Deposit** | Pick an amount from $10 to $100,000 and a coin: USDC, USDT or SOL on Solana; USDC, USDT or ETH on Ethereum; USDC, USDT or BNB on BNB Smart Chain; USDT or TRX on Tron; or BTC. Native coins go up to $20,000. The quote holds for 30 minutes, and the payment is credited as USDC on MLX Network. | | **Send** | To a name.mlx or a public wallet address. Instant and free. | | **Buy & sell** | The protocol sells MLX along its emission curve, at the curve price plus a premium, and buys it back at the curve price less a discount. The wallet shows both before you confirm. Sells depend on the treasury's USDC. | | **Withdraw** | USDC to a Solana address, at least 10 USDC plus a 1 USDC network fee. The amount waits in escrow on the ledger until the payout is sent by hand; the Solana signature is then recorded. A rejected withdrawal is refunded with its fee. | USDC on MLX is a claim on reserves the operator holds, not USDC issued by Circle. [Proof of reserves](https://mlx.app/reserves) shows the reserve wallets once they're published. #### Locked MLX Some MLX arrives locked. It counts toward your total, can't be sent or sold before its date, and unlocks by itself. - **Dedicated seats:** a seat pays nothing up front. It sets how long your mined MLX stays locked: 3 days for Ultra, 7 for Pro, 10 for Core, 14 for Lite and 20 for Nano. The smaller the seat, the longer the lock. - **Mined MLX:** each day's mining from your Macs is one lock that opens 30 days after that day on a free seat, or after your shortest seat lock-in with a dedicated seat. - **Before 20 October 2026:** dedicated seats were also paid a reward up front, and invites earned a bonus. Those stay locked until their original dates. MLX you buy at spot is never locked. Tap **Locked MLX** in the wallet to see every lock and when it opens. While your Macs work, that row counts up live. #### Activity, Macs and devices - **Activity** lists your transactions, with one **Mined on your Macs** row per day. Each token's page has its full history, filters and **Export CSV**, and **View in explorer** opens the public record. - **Macs** in Settings lists the Macs that work for this wallet. Linking one needs a signature from the Mac's own Secure Enclave key and one from your wallet, so no one can attach a Mac to your wallet, or yours to theirs, alone. A wallet can link up to 32 devices. - **Devices** lists the browsers where this wallet was opened, from mlx.app's visit records, so you can spot one you don't recognize. #### What MLX can and can't do with your funds From the [Terms](https://mlx.app/terms) and [risk disclosure](https://mlx.app/risks). MLX Network is centralized for now; the whitepaper describes the move to more sequencers. | The operator can't | The operator can | | --- | --- | | Sign a transaction from your account | Run the only sequencer: order transactions, pause the ledger or refuse a transaction | | See or recover your passkey, seed or recovery phrase | Correct the ledger after a bug or attack, explaining what it did | | Reverse a transfer you signed | Delay or refuse withdrawals and freeze balances for compliance checks | | Create new MLX: all of it was made in block 0 | Hold the reserves behind USDC on MLX and pay withdrawals by hand | That makes your key yours and your balances dependent on one operator. MLX Tokens can lose value or become worthless. Read the [risks](https://mlx.app/risks) before you deposit or buy anything. #### Questions **Where is my MLX wallet's private key stored?** Nowhere on MLX's servers. The seed it comes from is in your passkey, synced by your passkey manager, and in your recovery phrase. While the wallet is unlocked it sits in the browser tab's session memory. **What is a name.mlx address?** Your account's public name on the MLX ledger, like yourname.mlx. People send you MLX or USDC by typing it, and anyone can look it up on the explorer. **Is the MLX wallet self-custodial?** For your key, yes: only you can sign. But the ledger is run by one operator, and USDC on MLX is a claim on reserves the operator holds, so your balances depend on that operator. **How long do withdrawals take?** They're reviewed and paid by hand on Solana. The wallet says within 24 hours, usually sooner, but they can be delayed or paused, as the risk disclosure explains. **Why can't I send some of my MLX?** It's locked. Mined MLX unlocks on its date, shown on the wallet's Locked MLX page. Until then it counts toward your total but can't be sent or sold. ### Apple's Minions, explained URL: https://mlx.app/learn/apple-minions-idle-mac-gpu-cloud · Updated 2026-10-20 · MLX Network **The short version** - Minions pools the idle GPUs of Apple silicon Macs to run inference jobs, such as embeddings and LLM completions, without buying new GPUs ([Apple](https://machinelearning.apple.com/research/distributed-gpu-cloud)). - It's built for one organisation: a server splits jobs into subtasks, Macs pull them, and a watchdog hands a subtask out again when its Mac goes quiet. - The paper reports queue times falling from up to 10+ hours to minutes, but gives no device counts or speed benchmarks. - MLX Network is a separate, independent project: anyone's Mac, answers checked against other Macs, rewards in MLX Tokens. It isn't affiliated with Apple. In October 2026 six Apple researchers published Minions, a system that turns the idle GPUs of an organisation's Macs into a shared cloud for machine-learning inference. Here's what the paper describes, what it leaves out, and how it compares with MLX Network, an open network of Macs. #### What Minions is "Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices" is a three-page paper by Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter of Apple, posted on 9 October 2026 ([arXiv](https://arxiv.org/abs/2610.12970)) and listed on Apple Machine Learning Research ([Apple](https://machinelearning.apple.com/research/distributed-gpu-cloud)). Its starting point: inference jobs such as embedding generation and LLM completion usually wait for central GPU clusters that are expensive to grow and often oversubscribed, while the GPUs in Apple silicon Macs sit largely idle in everyday use. Minions asks whether those idle GPUs can be pooled into a shared inference platform "without additional hardware investment" ([paper](https://arxiv.org/pdf/2610.12970)). #### How it works - **Minion Server** takes jobs through a web UI and APIs, checks them and splits each one into subtasks. - **A PostgreSQL database** keeps every task, subtask and heartbeat. The Macs never touch it directly. - **Minion Executor** runs on each Mac: it pulls subtasks over REST, downloads the inputs from object storage (Amazon S3) and the model if it isn't cached, runs it on the Mac's GPU and writes the results back. - **Minion Watchdog** follows the heartbeats, every five minutes, of Macs that are running work, and puts a subtask back in the queue when its Mac sleeps, loses the network, powers off or is stopped by its owner. There's no central scheduler: Macs race to claim a subtask, and a unique constraint in the database lets exactly one win. Jobs run first in, first out, but a Mac's owner can point it at one team's project first. Embeddings run with Hugging Face transformers and LLM completions with Ollama, on the Mac's GPU ([paper](https://arxiv.org/pdf/2610.12970)). #### What the paper reports Minions "has been used to support GPU inference workloads across multiple teams". Its one table compares before and after: GPU queue time from hours (up to 10+) to minutes, new GPU requests from required to none, infrastructure cost from high to zero, and scaling effort from manual to organic ([paper](https://arxiv.org/pdf/2610.12970)). It also reports the human side. Team members ask colleagues over Slack to run the executor for a project, and some people worried about their data running on devices outside their own team. Minions answers that with deployments scoped to one team or organisation, each with its own database and pool of Macs. #### What it doesn't claim - **No benchmarks.** No tokens a second, no number of Macs, no comparison with data-centre GPUs. - **No training.** Distributed training is named as future work. - **No single giant GPU.** Each subtask runs on one Mac; the paper doesn't describe splitting one model across several Macs. - **No result checking.** The paper doesn't describe verifying the answers Macs send back. - **No pay and no product.** Participating Macs aren't rewarded, and the paper announces no Apple service. Its conclusion looks further ahead: if idle Macs "across organizations—or potentially across the global community—can be securely federated", they could form a large shared GPU pool. In the paper that's a direction for the future, not something it built ([paper](https://arxiv.org/pdf/2610.12970)). #### How MLX Network differs Sources: [the Minions paper](https://arxiv.org/pdf/2610.12970) and the [MLX Network whitepaper](https://mlx.app/whitepaper). | | Minions (Apple research) | MLX Network | | --- | --- | --- | | Who takes part | Macs inside one team or organisation | Anyone with an Apple Silicon Mac and [MLX Node](https://mlx.app/download) | | What runs | Embeddings (Hugging Face) and LLM completions (Ollama) | LLM jobs on Apple's MLX framework | | Getting work | Macs pull subtasks from a central server | MLX Node asks the network for jobs, within the limits you set | | Checking answers | Not described | Compared with other Macs' answers and known test jobs | | Reward | None: existing hardware at zero cost | MLX Tokens every hour, locked for 30 days on a free seat | | Status | Research paper; training is future work | Mainnet since 18 October 2026, every block on the public [explorer](https://explorer.mlx.app) | Both start from the same observation: Apple silicon Macs have capable GPUs that sit idle most of the day. Minions keeps that inside one organisation, among Macs it already trusts. MLX Network opens it to anyone, which is why it checks answers against other Macs and pays for verified work. MLX Network is independent and not affiliated with Apple; its name comes from the MLX framework it runs on ([why the name](https://mlx.app/learn/mlx-network-vs-apple-mlx)). Want your Mac to take part? [MLX Node](https://mlx.app/download) is free and works only when and as hard as you allow; the rules are in the [whitepaper](https://mlx.app/whitepaper), and every Mac on the network is on [MLX World](https://world.mlx.app). MLX Tokens can lose value: read the [risks](https://mlx.app/risks). #### Questions **Is Apple launching a GPU cloud made of Macs?** The paper doesn't say so. It describes a system teams already use for inference, and names training and federating idle Macs across organisations as future directions. It announces no product. **Does Minions combine several Macs into one GPU?** No. A job is split into independent subtasks and each one runs on a single Mac. The paper doesn't describe running one model across several Macs. **Which models does Minions run?** Embedding models from Hugging Face, through the transformers library, and LLM completions through Ollama, on the Macs' GPUs. The paper names no specific models. **Is MLX Network part of Apple or of Minions?** No. MLX Network is an independent network of Macs running MLX Node. It isn't affiliated with Apple and has nothing to do with the Minions project.