at-home/run-on/m4pro-48
What can I run on a MacBook Pro M4 Pro with 48 GB?
Of its 48 GB, about 37 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 Q6_K, decoding at up to 8 tokens a second.
| Model | Best fit | Download | Longest context | Decode, up to |
|---|---|---|---|---|
| Llama 3.2 3B | Q8_0 | 3.4 GB | 283K tokens | 64 tok/s |
| Qwen3.5 4B | Q8_0 | 4.3 GB | 964K tokens | 51 tok/s |
| Llama 3.1 8B | Q8_0 | 8.5 GB | 209K tokens | 26 tok/s |
| Qwen3 8B | Q8_0 | 8.7 GB | 185K tokens | 25 tok/s |
| Qwen3.5 9B | Q8_0 | 9.6 GB | 806K tokens | 23 tok/s |
| Qwen3 14B | Q8_0 | 15.7 GB | 124K tokens | 14 tok/s |
| gpt-oss 20B | MXFP4, as shipped | 12.8 GB | 946K tokens | 17 tok/s |
| Qwen3 32B | Q6_K | 26.9 GB | 36K tokens | 8 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 273 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.