Benchmarks · community-measured

Apple Silicon LLM benchmarks

Measured local-inference performance, contributed by rapid-mlx users: 22 model-and-chip combinations from 44 submissions across 6 machines (updated 2026-08-25). Explore and filter on the interactive leaderboard, or jump straight to a measurement below. Every number is reproducible: rapid-mlx bench <model> --tier speed --submit.

By model

`bonsai-1.7b-2bit`

Apple M5 Max329 tok/s
Apple M4 Max315 tok/s

`bonsai-27b-2bit`

Apple M23.8 tok/s

`gemma-4-26b-4bit`

Apple M4 Max108 tok/s

`gpt-oss-20b-mxfp4-q8`

Apple M5 Max107 tok/s
Apple M2 Pro44.6 tok/s

`qwen2.5-14b-4bit`

Apple M4 Max45.8 tok/s
Apple M5 Max43.5 tok/s

`qwen3.5-27b-4bit`

Apple M5 Max25.3 tok/s

`qwen3.5-4b-4bit`

Apple M4 Max138 tok/s
Apple M5 Max130 tok/s
Apple M2 Pro56.4 tok/s
Apple M437.3 tok/s
Apple M215.5 tok/s

`qwen3.5-9b-4bit`

Apple M5 Max78.9 tok/s
Apple M2 Pro34.6 tok/s

`qwen3.5-9b-8bit`

Apple M5 Max45.2 tok/s

`qwen3.6-27b-4bit`

Apple M4 Max25.9 tok/s

`qwen3.6-35b-4bit`

Apple M5 Pro86.3 tok/s

`qwen3.6-35b-optiq-4bit`

Apple M5 Max92.9 tok/s

`qwen3.8-27b-4bit`

Apple M5 Max25.3 tok/s

`qwen3.8-27b-mixed-3.5bpw`

Apple M5 Max26.5 tok/s

By chip

Apple M2

qwen3.5-4b-4bit15.5 tok/s
bonsai-27b-2bit3.8 tok/s

Apple M2 Pro

qwen3.5-4b-4bit56.4 tok/s
gpt-oss-20b-mxfp4-q844.6 tok/s
qwen3.5-9b-4bit34.6 tok/s

Apple M4

qwen3.5-4b-4bit37.3 tok/s

Apple M4 Max

bonsai-1.7b-2bit315 tok/s
qwen3.5-4b-4bit138 tok/s
gemma-4-26b-4bit108 tok/s
qwen2.5-14b-4bit45.8 tok/s
qwen3.6-27b-4bit25.9 tok/s

Apple M5 Max

Apple M5 Pro

qwen3.6-35b-4bit86.3 tok/s

Add your machine

$ curl -fsSL https://rapidmlx.com/install.sh | bash

$ rapid-mlx bench qwen3.5-9b-4bit --tier speed --submit

Raw corpus: /api/benchmarks · Submissions are anonymous and retractable — see the leaderboard for details.