at-home/run-on/m4-mini-32
What can I run on a Mac mini M4 with 32 GB?
Of its 32 GB, about 24 GB is left for a model's weights and cache, because macOS caps how much memory the GPU may pin and the engine needs some for itself. The largest model below that runs with room for a 8K conversation is Qwen3 32B at Q4_K_M, decoding at up to 5 tokens a second.
| Model | Best fit | Download | Longest context | Decode, up to |
|---|---|---|---|---|
| Llama 3.2 3B | Q8_0 | 3.4 GB | 175K tokens | 28 tok/s |
| Qwen3.5 4B | Q8_0 | 4.3 GB | 586K tokens | 23 tok/s |
| Llama 3.1 8B | Q8_0 | 8.5 GB | 115K tokens | 11 tok/s |
| Qwen3 8B | Q8_0 | 8.7 GB | 101K tokens | 11 tok/s |
| Qwen3.5 9B | Q8_0 | 9.6 GB | 427K tokens | 10 tok/s |
| Qwen3 14B | Q8_0 | 15.7 GB | 49K tokens | 6 tok/s |
| gpt-oss 20B | MXFP4, as shipped | 12.8 GB | 441K tokens | 8 tok/s |
| Qwen3 32B | Q4_K_M | 19.9 GB | 15K tokens | 5 tok/s |
| Llama 3.3 70B | does not fit |
Best fit is the highest-quality quantisation that still leaves room for a 8K context. It stops at Q8 rather than FP16, because Q8 is indistinguishable from FP16 in practice and reads half the bytes, so it decodes about twice as fast on a machine like this.
Decode is a ceiling: 80% of this machine’s 120 GB/s divided by the bytes a token reads, at the start of a conversation. On a measured 16 GB MacBook Air that arithmetic predicted 35.6 tokens a second and MLX delivered 31.4; llama.cpp delivered 17 on the same model. Expect your engine to land at or below it.
Longest context assumes an unquantised KV cache. Quantising the cache to Q8 roughly doubles it. Reading a long prompt also needs working memory the table does not count, so leave a margin.