at-home/run-on/m2-air-16
What can I run on a MacBook Air M2 with 16 GB?
Of its 16 GB, about 11 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 14B at Q4_K_M, decoding at up to 9 tokens a second.
| Model | Best fit | Download | Longest context | Decode, up to |
|---|---|---|---|---|
| Llama 3.2 3B | Q8_0 | 3.4 GB | 66K tokens | 24 tok/s |
| Qwen3.5 4B | Q8_0 | 4.3 GB | 207K tokens | 19 tok/s |
| Llama 3.1 8B | Q8_0 | 8.5 GB | 20K tokens | 9 tok/s |
| Qwen3 8B | Q8_0 | 8.7 GB | 16K tokens | 9 tok/s |
| Qwen3.5 9B | Q8_0 | 9.6 GB | 49K tokens | 8 tok/s |
| Qwen3 14B | Q4_K_M | 9.0 GB | 13K tokens | 9 tok/s |
| gpt-oss 20B | does not fit | |||
| Qwen3 32B | does not fit | |||
| 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 100 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.