M4 Pro
Apple M4 Pro: 24–64 GB of memory at 273 GB/s. Below: the open models it can run, estimated at 64 GB with an 8K context. Up to 64 GB of unified memory at 273 GB/s (64 GB in the Mac mini): it holds 70B models at 4-bit, but generation at that size is slow.
Last updated: 2026-10-14
Specifications
What it can run
Best-quality quantization that fits with a 8K context, and the estimated generation speed. Estimates, not benchmarks. See the methodology.
| Model | Params | Best quant | Memory | tok/s (est.) |
|---|---|---|---|---|
| Llama 3.3 70B Instruct | 70B | Q4_K_M | 44.5 GiB | 4.2 |
| Qwen3.5-35B-A3B | 35B (3B) | Q8_0 | 37.0 GiB | 56 |
| Qwen3.6-35B-A3B | 35B (3B) | Q8_0 | 37.0 GiB | 56 |
| Qwen3-32B | 32.8B | Q8_0 | 36.6 GiB | 5.1 |
| DeepSeek-R1-Distill-Qwen-32B | 32.76B | Q8_0 | 36.5 GiB | 5.1 |
| Gemma 4 31B IT | 30.7B | Q8_0 | 33.5 GiB | 5.4 |
| Qwen3-30B-A3B (2507) | 30.5B (3.3B) | Q8_0 | 32.9 GiB | 51 |
| Qwen3-Coder-30B-A3B-Instruct | 30.5B (3.3B) | Q8_0 | 32.9 GiB | 51 |
| GLM-4.7-Flash (30B-A3B) | 30B (3B) | Q8_0 | 32.1 GiB | 56 |
| Qwen3.5-27B | 27B | Q8_0 | 29.1 GiB | 6.2 |
| Qwen3.6-27B | 27B | Q8_0 | 29.1 GiB | 6.2 |
| Qwen3.8-27B | 27B | Q8_0 | 29.1 GiB | 6.2 |
| Gemma 3 27B IT | 27B | Q8_0 | 29.6 GiB | 6.2 |
| Gemma 4 26B A4B IT | 25.2B (3.8B) | Q8_0 | 27.0 GiB | 44 |
| Mistral Small 3.2 24B Instruct (2506) | 24B | Q8_0 | 26.7 GiB | 7.0 |
| Magistral Small 1.2 (2509) | 24B | Q8_0 | 26.7 GiB | 7.0 |
| Devstral Small 2 24B Instruct (2512) | 24B | Q8_0 | 26.7 GiB | 7.0 |
| gpt-oss-20b | 21B (3.6B) | FP16 | 41.8 GiB | 25 |
| Phi-4 (14B) | 14B | FP16 | 29.4 GiB | 6.3 |
| Gemma 3 12B IT | 12B | FP16 | 24.8 GiB | 7.4 |
| Gemma 4 12B Unified IT | 11.95B | FP16 | 24.2 GiB | 7.4 |
| Qwen3.5-9B | 9B | FP16 | 18.4 GiB | 9.9 |
| Qwen3-8B | 8.2B | FP16 | 17.7 GiB | 11 |
| DeepSeek-R1-0528-Qwen3-8B | 8.19B | FP16 | 17.6 GiB | 11 |
| Llama 3.1 8B Instruct | 8B | FP16 | 17.1 GiB | 11 |
| Gemma 4 E4B IT | 8B | FP16 | 16.4 GiB | 11 |
| Llama 3.2 3B Instruct | 3.21B | FP16 | 7.7 GiB | 28 |
| Kimi K2.6 | 1000B (32B) | Doesn't fit | – | – |
| DeepSeek-R1-0528 (671B) | 671B (37B) | Doesn't fit | – | – |
| DeepSeek-V4-Flash | 284B (13B) | Doesn't fit | – | – |
| Qwen3.8-Flash-Next | 180B (6B) | Doesn't fit | – | – |
| Qwen3.5-122B-A10B | 122B (10B) | Doesn't fit | – | – |
| Mistral Small 4 119B (2603) | 119B (6.5B) | Doesn't fit | – | – |
| gpt-oss-120b | 117B (5.1B) | Doesn't fit | – | – |
| Llama 4 Scout 17B-16E Instruct | 109B (17B) | Doesn't fit | – | – |
| GLM-4.5-Air | 106B (12B) | Doesn't fit | – | – |
| Qwen3-Coder-Next (80B-A3B) | 80B (3B) | Doesn't fit | – | – |
Prices
No prices recorded for this device yet. All prices →
Frequently asked questions
Which LLMs can the M4 Pro run?
With 64 GB and an 8K context, the largest models that fit include: Llama 3.3 70B Instruct, Qwen3.5-35B-A3B, Qwen3.6-35B-A3B, Qwen3-32B, DeepSeek-R1-Distill-Qwen-32B. See the full table below.
How fast is the M4 Pro for local LLMs?
Generation speed is mostly set by memory bandwidth (273 GB/s). For Llama 3.1 8B Instruct at FP16, expect roughly 11 tokens/s; bigger models are slower.
Can the M4 Pro run 70B models?
Yes. Llama 3.3 70B Instruct fits on the M4 Pro (64 GB) at Q4_K_M with an 8K context, using about 44.5 GiB, at an estimated 4.2 tokens/s.