Log inCreate wallet

The best LLMs for a 512 GB Mac.

61 of 73 open models run well in 512 GB. The GPU can use about 384 GB of it by default. Speed depends on the chip: from 819 GB/s on the M3 Ultra to 1,200 GB/s on the M5 Ultra.

Smartest that runs wellDeepSeek V4.1 Flash 10174-bit · 311 GB · 41–107 tok/sBiggest that runs wellDeepSeek V4.1 Flash 10174-bit · 311 GB · 41–107 tok/sFor codeQwen3 Coder Next8-bit · 85 GB · 115–300 tok/s

Macs with 512 GB

2 chips from M3 to M5 come with 512 GB. Open one for its own list and speeds.

Runs well 68

Fits with room for context, at 10 or more tokens a second on the M5 Ultra.

ModelVersionMemoryM3 UltraM5 Ultra
DeepSeek V4.1 Flash 1017552B · 16B active · Intelligence 45.14-bit311 GB41–7360–107
GLM-5.3 Flash320B · 18B active · Intelligence 41.88-bit340 GB19–3428–50
Qwen3.8 Flash Next180B · 6B active · Intelligence 39.88-bit191 GB58–10284–150
MiMo V2.6 Flash309B · 15B active · Intelligence 37.9native 4.48-bit173 GB44–7864–114
Qwen3.8 27B27B · Intelligence 33.78-bit28.7 GB13–2319–33
Beam 1.0501B · 23B active · Intelligence 30.34-bit282 GB29–5142–74
MiniMax M3428B · 23B active · Intelligence 29.26-bit348 GB20–3529–51
Inkling Small276B · 12B active · Intelligence 25.78-bit293 GB29–5142–75
Hy3295B · 21B active · Intelligence 25.38-bit313 GB17–2924–43
Nemotron 3 Ultra 550B A55B550B · 55B active · Intelligence 22.94-bit309 GB12–2117–31
Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.48-bit78.6 GB38–6856–100
Qwen3.5 397B A17B397B · 17B active · Intelligence 18.46-bit323 GB27–4739–70
Qwen3.6 35B A3B35B · 3B active · Intelligence 18.28-bit37.2 GB115–205169–300
Muse Glimmer 30B29.6B · Intelligence 17.58-bit31.5 GB12–2117–30
Qwen3.5 122B A10B122B · 10B active · Intelligence 15.68-bit130 GB35–6251–91
Gemma 4 31B30.7B · Intelligence 14.78-bit32.6 GB11–2017–29
Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.98-bit31.9 GB115–205169–300
Granite 4.2 30B30B · Intelligence 12.78-bit31.9 GB12–2117–30
MiniCPM5 2B2.52B · Intelligence 12.58-bit2.7 GB137–243200–356
gpt-oss-120b117B · 5.1B active · Intelligence 11.6native 4.25-bit65.3 GB137–243200–356
Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.38-bit126 GB53–9578–139
Qwen3.5 9B9B · Intelligence 11.28-bit9.6 GB38–6856–100
Granite 4.2 8B8B · Intelligence 11.18-bit8.5 GB43–7764–113
Llama 4 Maverick 17B-128E Instruct400B · 17B active · Intelligence 9.36-bit325 GB27–4739–70
Qwen3 Coder Next80B · 3B active · Intelligence 9.28-bit85 GB115–205169–300
Granite 4.2 3B3B · Intelligence 9.18-bit3.2 GB115–205169–300
gpt-oss-20b21B · 3.6B active · Intelligence 9native 4.25-bit13.8 GB194–345284–505
Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.98-bit31.9 GB100–177146–259
Gemma 4 E4B8B · Intelligence 8.98-bit8.5 GB77–137113–200
Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.58-bit116 GB20–3630–53
Ministral 3 8B8.9B · Intelligence 5.58-bit9.5 GB39–6957–101
Nemotron 3 Nano 4B3.97B · Intelligence 4.88-bit4.2 GB88–156129–229
DeepSeek V4 Flash 0731284B · 13B active8-bit302 GB27–4739–70
MiniMax M2.7230B · 10B active8-bit244 GB35–6251–91
Ornith 1.5 122B122B · 10B active8-bit130 GB35–6251–91
Nemotron 3 Super 120B A12B120B · 12B active8-bit128 GB29–5142–75
GLM-4.5 Air106B · 12B active8-bit113 GB29–5142–75
Llama 3.3 70B Instruct70B4-bit39.4 GB5–914–24
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active8-bit32.4 GB105–187154–274
GLM-4.7 Flash30B · 3B active8-bit31.9 GB115–205169–300
Qwen3.6 27B27B8-bit28.7 GB13–2319–33
Gemma 4 26B A4B25.2B · 3.8B active8-bit26.8 GB92–164135–240
Devstral Small 2 (24B)24B8-bit25.5 GB14–2621–38
Magistral Small 1.2 (2509)24B8-bit25.5 GB14–2621–38
Mistral Small 3.2 (24B)24B8-bit25.5 GB14–2621–38
LFM2.5 24B A2B24B · 2B active8-bit25.5 GB176–312257–457
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB23–4134–60
Phi-4 (14B)14B8-bit14.9 GB25–4436–64
Gemma 4 12B Unified12B8-bit12.7 GB29–5243–76
Qwen3 8B8.2B8-bit8.7 GB42–7562–110
Llama 3.1 8B Instruct8B8-bit8.5 GB43–7764–113
Gemma 4 E2B5.1B8-bit5.4 GB154–273225–400
Qwen3.5 4B4B8-bit4.3 GB86–152126–223
Phi-4 mini instruct3.8B8-bit4 GB92–164135–240
Llama 3.2 3B Instruct3.21B8-bit3.4 GB108–193159–282
SmolLM3 3B3B8-bit3.2 GB115–205169–300
LFM2.5 2.6B2.69B8-bit2.9 GB127–226186–331
Qwen3.5 2B2B8-bit2.1 GB176–312257–457
Llama 3.2 1B Instruct1.23B8-bit1.3 GB284–504415–738
Qwen3.5 0.8B0.8B8-bit0.9 GB410–728600–1067
Qwen3 0.6B0.6B8-bit0.6 GB614–1092900–1600
MiMo V2.6 Pro1,020B · 42B active · Intelligence 46.32-bit lower quality319 GB28–5041–73
GLM-5.3744B · 40B active · Intelligence 44.83-bit lower quality326 GB21–3731–55
Muse Spark Open816B · 42B active · Intelligence 38.92-bit lower quality255 GB28–5041–73
Kimi K2.7 Code1,000B · 32B active · Intelligence 25.82-bit lower quality313 GB37–6654–96
Inkling975B · 41B active · Intelligence 252-bit lower quality305 GB29–5142–75
Kimi K2.61,000B · 32B active2-bit lower quality313 GB37–6654–96
Mistral Large 3 (675B)675B · 41B active3-bit lower quality295 GB21–3730–54

Runs, but slowly 2

Fits, but under 10 tokens a second: fine for batch work, slow for chat.

ModelVersionMemoryM3 UltraM5 Ultra
Mistral Medium 3.5 (128B)128B · Intelligence 14.28-bit136 GB3–54–7
Devstral 2 (123B)123B8-bit131 GB3–54–7

Too big 3

Needs more memory than 512 GB, even at 2 bits.

Questions

What is the best LLM for a Mac with 512 GB?
Of the models with a published Intelligence Index, DeepSeek V4.1 Flash 1017 scores highest among those that run well in 512 GB (4-bit). For code: Qwen3 Coder Next.
Is 512 GB enough for local AI on a Mac?
61 of the 73 open models here run well in it at 4 bits or better. macOS lets the GPU use about 384 GB of it by default; the model and its context have to fit in that.
Can a 512 GB Mac run gpt-oss-20b?
Yes, native 4.25-bit, at about 194–345 tokens a second on an M3 Ultra and 284–505 on an M5 Ultra.

Other sizes: 8 GB · 16 GB · 18 GB · 24 GB · 32 GB · 36 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB

Run them with MLX Harness (free, open source) or Apple’s mlx-lm in Terminal: step by step.

How it’s worked out

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 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. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.

Best local LLMs for a 512 GB Mac (M3 to M5): what runs, how fast