Benchmark result
mlx-community/Qwen3.8-27B-4bit on Apple M5 Max — 89.7 tok/s
Measured with splash manual on November 5, 2026.
Support llm-bench.io
llm-bench.io is free. If it helped you, I'd appreciate your support to cover the infrastructure costs.
What the model built
Open full screen →The coding scenario asks for a playable game in a single HTML file. This is exactly what the model returned, unedited; the frame only adds a line that tells this page how tall it is.
How this run compares
4th fastest of 17 runs of this model on Apple M5 Max · median 31.9 tok/s.
sort
- 199.4tok/s75.8splash
- 297.6tok/s81.9splash
- 390.5tok/s85.2vLLM
- 489.7tok/s88.8splashthis run
- 584.4tok/s86.9OpenAI-compatible
- 684.0tok/s87.7vLLM
- 782.2tok/s82.6vLLM
- 874.9tok/s83.0vLLM
- 931.9tok/s76.0oMLX
- 1031.3tok/s72.4oMLX
- 1130.8tok/s78.2oMLX
- 1230.6tok/s84.6oMLX
- 1330.5tok/s84.2oMLX
- 1430.5tok/s87.1oMLX
- 1530.4tok/s83.8oMLX
- 1630.2tok/s84.5oMLX
- 1729.2tok/s85.7oMLX
mlx-community/Qwen3.8-27B-4bit on other hardware
Reproduce this run
toolsplash manual
modelmlx-community/Qwen3.8-27B-4bit
context65,536
thinkingon
effortxhigh
clientv0.4.98+102
Set those in splash, then:
llm-benchmark benchmark --model "mlx-community/Qwen3.8-27B-4bit" --tool "splash"The client prompts for context length and thinking mode, and for the KV cache dtype on Unsloth Studio. Sampling is left at the model default — the values above are what the tool reported using, not overrides the client sent.