Runs well 21
Fits with room for context, at 10 or more tokens a second.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| MiniCPM5 2B2.52B · Intelligence 12.5 | 8-bit | 2.7 GB | 17–30 |
| Granite 4.2 8B8B · Intelligence 11.1 | 4-bit | 4.5 GB | 10–18 |
| Granite 4.2 3B3B · Intelligence 9.1 | 8-bit | 3.2 GB | 14–25 |
| Gemma 4 E4B8B · Intelligence 8.9 | 4-bit | 4.5 GB | 18–32 |
| Nemotron 3 Nano 4B3.97B · Intelligence 4.8 | 8-bit | 4.2 GB | 11–19 |
| Llama 3.1 8B Instruct8B | 4-bit | 4.5 GB | 10–18 |
| Gemma 4 E2B5.1B | 6-bit | 4.1 GB | 24–42 |
| Qwen3.5 4B4B | 8-bit | 4.3 GB | 10–19 |
| Phi-4 mini instruct3.8B | 8-bit | 4 GB | 11–20 |
| Llama 3.2 3B Instruct3.21B | 8-bit | 3.4 GB | 13–24 |
| SmolLM3 3B3B | 8-bit | 3.2 GB | 14–25 |
| LFM2.5 2.6B2.69B | 8-bit | 2.9 GB | 16–28 |
| Qwen3.5 2B2B | 8-bit | 2.1 GB | 21–38 |
| Llama 3.2 1B Instruct1.23B | 8-bit | 1.3 GB | 35–62 |
| Qwen3.5 0.8B0.8B | 8-bit | 0.9 GB | 50–89 |
| Qwen3 0.6B0.6B | 8-bit | 0.6 GB | 75–133 |
| Qwen3.5 9B9B · Intelligence 11.2 | 3-bit lower quality | 3.9 GB | 12–21 |
| Ministral 3 8B8.9B · Intelligence 5.5 | 3-bit lower quality | 3.9 GB | 12–21 |
| Phi-4 (14B)14B | 2-bit lower quality | 4.4 GB | 10–18 |
| Gemma 4 12B Unified12B | 2-bit lower quality | 3.7 GB | 12–22 |
| Qwen3 8B8.2B | 3-bit lower quality | 3.6 GB | 13–22 |
Only just fits 1
Needs almost all memory: raise the GPU memory limit and keep the context short.
| Model | Version | Memory | Tokens/s |
|---|---|---|---|
| Phi-4 Reasoning Vision 15B15B | 2-bit lower quality | 4.7 GB | 10–17 |
Too big 51
Needs more memory than 8 GB, even at 2 bits.
- MiMo V2.6 Pro1,020B
- Kimi K32,800B
- MiMo V2.6 Flash309B
- Kimi K2.7 Code1,000B
- gpt-oss-120b117B
- gpt-oss-20b21B
- Kimi K2.61,000B
- DeepSeek V4.1 Flash 1017552B
- GLM-5.3744B
- GLM-5.3 Flash320B
- Qwen3.8 2.4T A95B2,400B
- Qwen3.8 Flash Next180B
- Muse Spark Open816B
- DeepSeek V4 Pro 08131,600B
- Qwen3.8 27B27B
- Beam 1.0501B
- MiniMax M3428B
- Inkling Small276B
- Hy3295B
- Inkling975B
- Nemotron 3 Ultra 550B A55B550B
- Nemotron 3.5 Bolt 74B A9B74B
- Qwen3.5 397B A17B397B
- Qwen3.6 35B A3B35B
- Muse Glimmer 30B29.6B
- Qwen3.5 122B A10B122B
- Gemma 4 31B30.7B
- Mistral Medium 3.5 (128B)128B
- Nemotron 3.5 Lightning 30B A3B30B
- Granite 4.2 30B30B
- Mistral Small 4 (119B)119B
- Llama 4 Maverick 17B-128E Instruct400B
- Qwen3 Coder Next80B
- Nemotron 3 Nano 30B A3B30B
- Llama 4 Scout 17B-16E Instruct109B
- Mistral Large 3 (675B)675B
- DeepSeek V4 Flash 0731284B
- MiniMax M2.7230B
- Devstral 2 (123B)123B
- Ornith 1.5 122B122B
- Nemotron 3 Super 120B A12B120B
- GLM-4.5 Air106B
- Llama 3.3 70B Instruct70B
- Qwen3 Coder 30B A3B Instruct30.5B
- GLM-4.7 Flash30B
- Qwen3.6 27B27B
- Gemma 4 26B A4B25.2B
- Devstral Small 2 (24B)24B
- Magistral Small 1.2 (2509)24B
- Mistral Small 3.2 (24B)24B
- LFM2.5 24B A2B24B
Other Macs
Every Mac with 8 GB, from any chip.
The M2 with more or less memory:
Neighbouring chips:
Head to head:
Compare it with any Mac or GPU · The M2 for local AI · The best Mac for local LLMs · Every Mac
Questions
- What is the best LLM for an M2 with 8 GB?
- Of the models with a published Intelligence Index, MiniCPM5 2B scores highest among those that run well here: 8-bit, about 17–30 tokens a second. The biggest that runs well is Granite 4.2 8B.
- Can an M2 with 8 GB run Llama 3.3 70B?
- No: even at 2 bits it needs more than the 6 GB the GPU can use by default.
- How fast is gpt-oss-20b on an M2?
- It doesn’t fit in 8 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.