llama.cpp on the M3 Ultra 512GB
On the M3 Ultra 512GB, llama.cpp runs Qwen3.8 125B · 6B active at 4-bit at a median 31.1 tokens per second for generation, from 1 plain run (250K context).
Speed by model on the M3 Ultra 512GB
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 |
|---|---|---|
| Qwen3.8 125B · 6B active 4-bit | 31.1 t/sone report, 250K context | oMLX 25.0 t/sone report, context not stated |
| Qwen3.8 125B · 6B active 8-bit | 25.7 t/sone report, 32K context | no other engine reported |
8 llama.cpp reports on the M3 Ultra 512GB
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.
- Qwen3.5 · MXFP4_MOE80.9 t/s
- Gemma 4 · MXFP4_MOE87.8 t/s
- Gemma 4 · UD-Q8_K_XL84.9 t/s
- Gemma 4 · UD-Q5_K_M88.1 t/s
- Gemma 4 · BF1662.4 t/s
- Qwen3.8 125B · 6B active · Q8_025.7 t/s
- Qwen3.8 125B · 6B active · Q4_K_XL31.1 t/s
- DeepSeek V4 Flash 284B · 13B active · Q4_K_XL8.0 t/s
Frequently asked
How fast is llama.cpp on the M3 Ultra 512GB?
On the M3 Ultra 512GB, llama.cpp runs Qwen3.8 125B · 6B active at 4-bit at a median 31.1 tokens per second for generation, from 1 plain run (250K context).
Is llama.cpp faster than other engines on the M3 Ultra 512GB?
Same card, same model size and bit class, plain runs only: Qwen3.8 125B · 6B active at 4-bit gives llama.cpp 31.1 t/s (1 run) and oMLX 25.0 t/s (1). Context length and engine build differ between the runs, so treat it as a lead, not a benchmark.