Benchmarks · community-measured
qwen3.5-4b-4bit on Apple M2
Real measured local-inference performance for
qwen3.5-4b-4bit on an Apple M2 (24 GB unified memory),
from the rapid-mlx community benchmark corpus. Latest submission 2026-08-15,
measured on rapid-mlx 0.12.12.
15.5
tok/s decode
93.1
tok/s prefill
23.1 s
first token
4.4 GB
peak memory
A single community run so far — treat it as one data point, not a distribution. Reproduce it below and submit yours to firm it up.
All measurements for this pair
| RAM | Accel | Decode | Prefill | TTFT | Peak mem | Runs | rapid-mlx | Latest |
|---|---|---|---|---|---|---|---|---|
| 24 GB | — | 15.5 tok/s | 93.1 tok/s | 23.1 s | 4.4 GB | 1 | 0.12.12 | 2026-08-15 |
Standardized single-request benchmark: 5 measured rounds after 1 discarded warmup, greedy sampling, prefix cache disabled. Raw corpus: /api/benchmarks.
Reproduce this number
Install rapid-mlx, run the same suite on your Mac, and submit the result — it lands on the community leaderboard and updates this page.
$ curl -fsSL https://rapidmlx.com/install.sh | bash
$ rapid-mlx bench qwen3.5-4b-4bit --tier speed --submit
More
qwen3.5-4b-4bit model page · Qwen family · Interactive leaderboard · Alias catalog