Runs well 48
Fits with room for context, at 10 or more tokens a second.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| Qwen3.8 27B27B · Intelligence 33.7 | 4-bit | 15.2 GB | 12–21 |
| Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.4 | 6-bit | 60.1 GB | 25–44 |
| Qwen3.6 35B A3B35B · 3B active · Intelligence 18.2 | 8-bit | 37.2 GB | 56–100 |
| Muse Glimmer 30B29.6B · Intelligence 17.5 | 4-bit | 16.7 GB | 11–19 |
| Gemma 4 31B30.7B · Intelligence 14.7 | 4-bit | 17.3 GB | 10–18 |
| Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.9 | 8-bit | 31.9 GB | 56–100 |
| Granite 4.2 30B30B · Intelligence 12.7 | 4-bit | 16.9 GB | 11–19 |
| MiniCPM5 2B2.52B · Intelligence 12.5 | 8-bit | 2.7 GB | 67–119 |
| gpt-oss-120b117B · 5.1B active · Intelligence 11.6 | native 4.25-bit | 65.3 GB | 67–119 |
| Qwen3.5 9B9B · Intelligence 11.2 | 8-bit | 9.6 GB | 19–33 |
| Granite 4.2 8B8B · Intelligence 11.1 | 8-bit | 8.5 GB | 21–38 |
| Qwen3 Coder Next80B · 3B active · Intelligence 9.2 | 6-bit | 65 GB | 75–133 |
| Granite 4.2 3B3B · Intelligence 9.1 | 8-bit | 3.2 GB | 56–100 |
| gpt-oss-20b21B · 3.6B active · Intelligence 9 | native 4.25-bit | 13.8 GB | 95–168 |
| Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.9 | 8-bit | 31.9 GB | 49–86 |
| Gemma 4 E4B8B · Intelligence 8.9 | 8-bit | 8.5 GB | 38–67 |
| Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.5 | 4-bit | 61.3 GB | 19–33 |
| Ministral 3 8B8.9B · Intelligence 5.5 | 8-bit | 9.5 GB | 19–34 |
| Nemotron 3 Nano 4B3.97B · Intelligence 4.8 | 8-bit | 4.2 GB | 43–76 |
| GLM-4.5 Air106B · 12B active | 4-bit | 59.6 GB | 26–47 |
| Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active | 8-bit | 32.4 GB | 51–91 |
| GLM-4.7 Flash30B · 3B active | 8-bit | 31.9 GB | 56–100 |
| Qwen3.6 27B27B | 4-bit | 15.2 GB | 12–21 |
| Gemma 4 26B A4B25.2B · 3.8B active | 8-bit | 26.8 GB | 45–80 |
| Devstral Small 2 (24B)24B | 4-bit | 13.5 GB | 13–24 |
| Magistral Small 1.2 (2509)24B | 4-bit | 13.5 GB | 13–24 |
| Mistral Small 3.2 (24B)24B | 4-bit | 13.5 GB | 13–24 |
| LFM2.5 24B A2B24B · 2B active | 8-bit | 25.5 GB | 86–152 |
| Phi-4 Reasoning Vision 15B15B | 8-bit | 15.9 GB | 11–20 |
| Phi-4 (14B)14B | 8-bit | 14.9 GB | 12–21 |
| Gemma 4 12B Unified12B | 8-bit | 12.7 GB | 14–25 |
| Qwen3 8B8.2B | 8-bit | 8.7 GB | 21–37 |
| Llama 3.1 8B Instruct8B | 8-bit | 8.5 GB | 21–38 |
| Gemma 4 E2B5.1B | 8-bit | 5.4 GB | 75–133 |
| Qwen3.5 4B4B | 8-bit | 4.3 GB | 42–74 |
| Phi-4 mini instruct3.8B | 8-bit | 4 GB | 45–80 |
| Llama 3.2 3B Instruct3.21B | 8-bit | 3.4 GB | 53–94 |
| SmolLM3 3B3B | 8-bit | 3.2 GB | 56–100 |
| LFM2.5 2.6B2.69B | 8-bit | 2.9 GB | 62–110 |
| Qwen3.5 2B2B | 8-bit | 2.1 GB | 86–152 |
| Llama 3.2 1B Instruct1.23B | 8-bit | 1.3 GB | 138–246 |
| Qwen3.5 0.8B0.8B | 8-bit | 0.9 GB | 200–356 |
| Qwen3 0.6B0.6B | 8-bit | 0.6 GB | 300–533 |
| Qwen3.8 Flash Next180B · 6B active · Intelligence 39.8 | 2-bit lower quality | 56.3 GB | 95–168 |
| Qwen3.5 122B A10B122B · 10B active · Intelligence 15.6 | 3-bit lower quality | 53.4 GB | 41–73 |
| Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.3 | 3-bit lower quality | 52.1 GB | 64–114 |
| Ornith 1.5 122B122B · 10B active | 3-bit lower quality | 53.4 GB | 41–73 |
| Nemotron 3 Super 120B A12B120B · 12B active | 3-bit lower quality | 52.5 GB | 34–60 |
Runs, but slowly 3
Fits, but under 10 tokens a second: fine for batch work, slow for chat.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| Llama 3.3 70B Instruct70B | 6-bit | 56.9 GB | 3–6 |
| Mistral Medium 3.5 (128B)128B · Intelligence 14.2 | 3-bit lower quality | 56 GB | 3–6 |
| Devstral 2 (123B)123B | 3-bit lower quality | 53.8 GB | 3–6 |
Only just fits 1
Needs almost all memory: raise the GPU memory limit and keep the context short.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| MiniMax M2.7230B · 10B active | 2-bit lower quality | 71.9 GB | 58–103 |
Too big 21
Needs more memory than 96 GB, even at 2 bits.
- MiMo V2.6 Pro1,020B
- Kimi K32,800B
- MiMo V2.6 Flash309B
- Kimi K2.7 Code1,000B
- Kimi K2.61,000B
- DeepSeek V4.1 Flash 1017552B
- GLM-5.3744B
- GLM-5.3 Flash320B
- Qwen3.8 2.4T A95B2,400B
- Muse Spark Open816B
- DeepSeek V4 Pro 08131,600B
- Beam 1.0501B
- MiniMax M3428B
- Inkling Small276B
- Hy3295B
- Inkling975B
- Nemotron 3 Ultra 550B A55B550B
- Qwen3.5 397B A17B397B
- Llama 4 Maverick 17B-128E Instruct400B
- Mistral Large 3 (675B)675B
- DeepSeek V4 Flash 0731284B
Other Macs
Every Mac with 96 GB, from any chip.
The M2 Max with more or less memory:
Neighbouring chips:
Head to head:
Compare it with any Mac or GPU · The M2 Max for local AI · The best Mac for local LLMs · Every Mac
Questions
- What is the best LLM for an M2 Max with 96 GB?
- Of the models with a published Intelligence Index, Qwen3.8 27B scores highest among those that run well here: 4-bit, about 12–21 tokens a second. The biggest that runs well is gpt-oss-120b.
- Can an M2 Max with 96 GB run Llama 3.3 70B?
- Yes, but slowly: 6-bit, at about 3–6 tokens a second.
- How fast is gpt-oss-20b on an M2 Max?
- About 95–168 tokens a second (native 4.25-bit), from the chip’s 400 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.