llama.cpp on the M5 Max 128GB
On the M5 Max 128GB, llama.cpp runs Qwen3.6 35B · 3B active at 8-bit at a median 79.4 tokens per second for generation, from 1 plain run (125K context).
Speed by model on the M5 Max 128GB
Plain runs only: one device, one request, no speculative decoding, the whole model in memory. Each row is one model size at one bit class, with the other engines' plain runs of the same model on this card beside it. Context length and build still differ between runs, and each side shows its own.
| Model | llama.cpp | Other engines |
|---|---|---|
| Qwen3.6 35B · 3B active 8-bit | 79.4 t/sone report, 125K context | no other engine reported |
| StepFun 3.7 198B · 11B active 4-bit | 33.9 t/sone report, 64K context | no other engine reported |
9 llama.cpp reports on the M5 Max 128GB
Newest first. A run on several cards, with speculative decoding, with part of the model in system RAM or with batched requests says so, since each describes a different setup.
- Qwen3.8 122B · 10B active · Q8_020.4 t/s(speculative decoding, part in system RAM)
- Kimi K3 2800B · 104B active · 4-bit1.0 t/s(part in system RAM)
- Qwen3.6 35B · 3B active · Q8_0105.0 t/s(speculative decoding)
- GLM-5.3 320B · 18B active · IQ1_S24.0 t/s(2 cards)
- Qwen3.6 35B · 3B active · Q8 XL79.4 t/s
- Qwen3.8 27B31.8 t/s(speculative decoding)
- Qwen3.8 125B · 6B active · UD-Q2_K_XL11.5 t/s(speculative decoding)
- StepFun 3.7 198B · 11B active · Q4_K_S33.9 t/s
- Tencent-HY3 295B · 21B active · UD12832.4 t/s
Frequently asked
How fast is llama.cpp on the M5 Max 128GB?
On the M5 Max 128GB, llama.cpp runs Qwen3.6 35B · 3B active at 8-bit at a median 79.4 tokens per second for generation, from 1 plain run (125K context).