Home › Models › Qwen3 4B Instruct 2507 › Apple M4 Max (40-core GPU)
Qwen3 4B Instruct 2507 on the Apple M4 Max (40-core GPU)
Qwen3 4B Instruct 2507 runs at about 135 tokens/s on the Apple M4 Max (40-core GPU) (measured, MLX-4bit, MLX, mlx_lm.benchmark).
Best measured generation135tok/s
MLX-4bit · MLX
Prompt processing1,622tok/s
same run
Every run
| Quant | Backend | Generate tok/s | Prompt tok/s | Test | Build / by / notes | Source |
|---|---|---|---|---|---|---|
| MLX-4bit | MLX | 135 | 1,622 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-4bit; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-4bit-g32 | MLX | 126 | 1,610 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-4bit-g32; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-5bit | MLX | 116 | 1,584 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-5bit; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-5bit-g32 | MLX | 110 | 1,571 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-5bit-g32; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-6bit | MLX | 105 | 1,577 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-6bit; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-8bit | MLX | 86.9 | 1,607 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights mlx-community/Qwen3-4B-Instruct-2507-8bit; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
| MLX-BF16 | MLX | 52.5 | 1,781 | tg128 | mlx-lm 0.28.2 mlx-lm maintainers 64 GB M4 Max (only sold with the 40-core GPU); weights Qwen/Qwen3-4B-Instruct-2507; size = its safetensors files | ml-explore/mlx-lm BENCHMARKS.md (MIT) |
tg128 = generating 128 tokens; pp512 = reading a 512-token prompt (llama-bench). LocalScore averages several prompt and output lengths, so its numbers run lower than tg128. "decode, llama-server" = single-stream generation speed measured against a running llama-server (no speculative decoding).
No card yet?
Try it on a rented GPU (RunPod)Affiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.
All results for the Apple M4 Max (40-core GPU) · Qwen3 4B Instruct 2507 on other hardware
Embed this result
HTML
<a href="https://rigbench.pages.dev/model/qwen3-4b-instruct-2507/on/apple-m4-max-40c/"><img src="https://rigbench.pages.dev/badge/qwen3-4b-instruct-2507--apple-m4-max-40c.svg" alt="Qwen3 4B Instruct 2507 · Apple M4 Max (40-core GPU): 135 tok/s (MLX-4bit, measured)"></a>
Markdown (GitHub README, forums)
[](https://rigbench.pages.dev/model/qwen3-4b-instruct-2507/on/apple-m4-max-40c/)