Qwen3.8 27B on the Apple M3 Ultra
Qwen3.8 27B runs at about 22.8 tokens/s on the Apple M3 Ultra (measured, Q6_K, Metal, single-stream llama-server decode).
Best measured: 22.8 tokens/s generation (Q6_K, Metal). The source gives no prompt-processing figure.
Every run
| Quant | Backend | Generate tok/s | Prompt tok/s | Test | Build / by / notes | Source |
|---|---|---|---|---|---|---|
| Q6_K | Metal | 22.8 | – | decode, llama-server, MTP off | Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. GPU core count not stated. | qwen38-mtp community table (GitHub) |
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 M3 Ultra · Qwen3.8 27B on other hardware
Embed this result
HTML
<a href="https://rigbench.pages.dev/model/qwen3-8-27b/on/apple-m3-ultra/"><img src="https://rigbench.pages.dev/badge/qwen3-8-27b--apple-m3-ultra.svg" alt="Qwen3.8 27B · Apple M3 Ultra: 22.8 tok/s (Q6_K, measured)"></a>
Markdown (GitHub README, forums)
[](https://rigbench.pages.dev/model/qwen3-8-27b/on/apple-m3-ultra/)