Runs well 68
Fits with room for context, at 10 or more tokens a second.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017552B · 16B active · Intelligence 45.1 | 4-bit | 311 GB | 60–107 |
| GLM-5.3 Flash320B · 18B active · Intelligence 41.8 | 8-bit | 340 GB | 28–50 |
| Qwen3.8 Flash Next180B · 6B active · Intelligence 39.8 | 8-bit | 191 GB | 84–150 |
| MiMo V2.6 Flash309B · 15B active · Intelligence 37.9 | native 4.48-bit | 173 GB | 64–114 |
| Qwen3.8 27B27B · Intelligence 33.7 | 8-bit | 28.7 GB | 19–33 |
| Beam 1.0501B · 23B active · Intelligence 30.3 | 4-bit | 282 GB | 42–74 |
| MiniMax M3428B · 23B active · Intelligence 29.2 | 6-bit | 348 GB | 29–51 |
| Inkling Small276B · 12B active · Intelligence 25.7 | 8-bit | 293 GB | 42–75 |
| Hy3295B · 21B active · Intelligence 25.3 | 8-bit | 313 GB | 24–43 |
| Nemotron 3 Ultra 550B A55B550B · 55B active · Intelligence 22.9 | 4-bit | 309 GB | 17–31 |
| Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.4 | 8-bit | 78.6 GB | 56–100 |
| Qwen3.5 397B A17B397B · 17B active · Intelligence 18.4 | 6-bit | 323 GB | 39–70 |
| Qwen3.6 35B A3B35B · 3B active · Intelligence 18.2 | 8-bit | 37.2 GB | 169–300 |
| Muse Glimmer 30B29.6B · Intelligence 17.5 | 8-bit | 31.5 GB | 17–30 |
| Qwen3.5 122B A10B122B · 10B active · Intelligence 15.6 | 8-bit | 130 GB | 51–91 |
| Gemma 4 31B30.7B · Intelligence 14.7 | 8-bit | 32.6 GB | 17–29 |
| Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.9 | 8-bit | 31.9 GB | 169–300 |
| Granite 4.2 30B30B · Intelligence 12.7 | 8-bit | 31.9 GB | 17–30 |
| MiniCPM5 2B2.52B · Intelligence 12.5 | 8-bit | 2.7 GB | 200–356 |
| gpt-oss-120b117B · 5.1B active · Intelligence 11.6 | native 4.25-bit | 65.3 GB | 200–356 |
| Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.3 | 8-bit | 126 GB | 78–139 |
| Qwen3.5 9B9B · Intelligence 11.2 | 8-bit | 9.6 GB | 56–100 |
| Granite 4.2 8B8B · Intelligence 11.1 | 8-bit | 8.5 GB | 64–113 |
| Llama 4 Maverick 17B-128E Instruct400B · 17B active · Intelligence 9.3 | 6-bit | 325 GB | 39–70 |
| Qwen3 Coder Next80B · 3B active · Intelligence 9.2 | 8-bit | 85 GB | 169–300 |
| Granite 4.2 3B3B · Intelligence 9.1 | 8-bit | 3.2 GB | 169–300 |
| gpt-oss-20b21B · 3.6B active · Intelligence 9 | native 4.25-bit | 13.8 GB | 284–505 |
| Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.9 | 8-bit | 31.9 GB | 146–259 |
| Gemma 4 E4B8B · Intelligence 8.9 | 8-bit | 8.5 GB | 113–200 |
| Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.5 | 8-bit | 116 GB | 30–53 |
| Ministral 3 8B8.9B · Intelligence 5.5 | 8-bit | 9.5 GB | 57–101 |
| Nemotron 3 Nano 4B3.97B · Intelligence 4.8 | 8-bit | 4.2 GB | 129–229 |
| DeepSeek V4 Flash 0731284B · 13B active | 8-bit | 302 GB | 39–70 |
| MiniMax M2.7230B · 10B active | 8-bit | 244 GB | 51–91 |
| Ornith 1.5 122B122B · 10B active | 8-bit | 130 GB | 51–91 |
| Nemotron 3 Super 120B A12B120B · 12B active | 8-bit | 128 GB | 42–75 |
| GLM-4.5 Air106B · 12B active | 8-bit | 113 GB | 42–75 |
| Llama 3.3 70B Instruct70B | 4-bit | 39.4 GB | 14–24 |
| Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active | 8-bit | 32.4 GB | 154–274 |
| GLM-4.7 Flash30B · 3B active | 8-bit | 31.9 GB | 169–300 |
| Qwen3.6 27B27B | 8-bit | 28.7 GB | 19–33 |
| Gemma 4 26B A4B25.2B · 3.8B active | 8-bit | 26.8 GB | 135–240 |
| Devstral Small 2 (24B)24B | 8-bit | 25.5 GB | 21–38 |
| Magistral Small 1.2 (2509)24B | 8-bit | 25.5 GB | 21–38 |
| Mistral Small 3.2 (24B)24B | 8-bit | 25.5 GB | 21–38 |
| LFM2.5 24B A2B24B · 2B active | 8-bit | 25.5 GB | 257–457 |
| Phi-4 Reasoning Vision 15B15B | 8-bit | 15.9 GB | 34–60 |
| Phi-4 (14B)14B | 8-bit | 14.9 GB | 36–64 |
| Gemma 4 12B Unified12B | 8-bit | 12.7 GB | 43–76 |
| Qwen3 8B8.2B | 8-bit | 8.7 GB | 62–110 |
| Llama 3.1 8B Instruct8B | 8-bit | 8.5 GB | 64–113 |
| Gemma 4 E2B5.1B | 8-bit | 5.4 GB | 225–400 |
| Qwen3.5 4B4B | 8-bit | 4.3 GB | 126–223 |
| Phi-4 mini instruct3.8B | 8-bit | 4 GB | 135–240 |
| Llama 3.2 3B Instruct3.21B | 8-bit | 3.4 GB | 159–282 |
| SmolLM3 3B3B | 8-bit | 3.2 GB | 169–300 |
| LFM2.5 2.6B2.69B | 8-bit | 2.9 GB | 186–331 |
| Qwen3.5 2B2B | 8-bit | 2.1 GB | 257–457 |
| Llama 3.2 1B Instruct1.23B | 8-bit | 1.3 GB | 415–738 |
| Qwen3.5 0.8B0.8B | 8-bit | 0.9 GB | 600–1067 |
| Qwen3 0.6B0.6B | 8-bit | 0.6 GB | 900–1600 |
| MiMo V2.6 Pro1,020B · 42B active · Intelligence 46.3 | 2-bit lower quality | 319 GB | 41–73 |
| GLM-5.3744B · 40B active · Intelligence 44.8 | 3-bit lower quality | 326 GB | 31–55 |
| Muse Spark Open816B · 42B active · Intelligence 38.9 | 2-bit lower quality | 255 GB | 41–73 |
| Kimi K2.7 Code1,000B · 32B active · Intelligence 25.8 | 2-bit lower quality | 313 GB | 54–96 |
| Inkling975B · 41B active · Intelligence 25 | 2-bit lower quality | 305 GB | 42–75 |
| Kimi K2.61,000B · 32B active | 2-bit lower quality | 313 GB | 54–96 |
| Mistral Large 3 (675B)675B · 41B active | 3-bit lower quality | 295 GB | 30–54 |
Runs, but slowly 2
Fits, but under 10 tokens a second: fine for batch work, slow for chat.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| Mistral Medium 3.5 (128B)128B · Intelligence 14.2 | 8-bit | 136 GB | 4–7 |
| Devstral 2 (123B)123B | 8-bit | 131 GB | 4–7 |
Too big 3
Needs more memory than 512 GB, even at 2 bits.
Other Macs
Every Mac with 512 GB, from any chip.
The M5 Ultra with more or less memory:
Neighbouring chips:
Head to head:
Against PCs and graphics cards:
Compare it with any Mac or GPU · The M5 Ultra for local AI · The best Mac for local LLMs · Every Mac
Questions
- What is the best LLM for an M5 Ultra with 512 GB?
- Of the models with a published Intelligence Index, DeepSeek V4.1 Flash 1017 scores highest among those that run well here: 4-bit, about 60–107 tokens a second.
- Can an M5 Ultra with 512 GB run Llama 3.3 70B?
- Yes: 4-bit, at about 14–24 tokens a second.
- How fast is gpt-oss-20b on an M5 Ultra?
- About 284–505 tokens a second (native 4.25-bit), from the chip’s 1,200 GB/s memory bandwidth.
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.