llama.cpp on the M4 Max 128GB
On the M4 Max 128GB, llama.cpp runs Qwen3.8 125B · 6B active at 3-bit at a median 27.3 tokens per second for generation, from 1 plain run (32K context).
Speed by model on the M4 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 |
|---|---|---|
| Qwen3.8 125B · 6B active 3-bit | 27.3 t/sone report, 32K context | no other engine reported |
33 llama.cpp reports on the M4 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 · Q8_015.7 t/s
- Qwen3.5 · Q8_048.4 t/s
- Qwen3.5 · Q8_078.8 t/s
- SmolLM3 · Q8_0119.7 t/s
- SmolLM3 · Q4_K_M168.6 t/s
- Qwen3.8 · UD-Q6_K_XL17.2 t/s
- Qwen3.8 · UD-Q5_K_XL19.7 t/s
- Qwen3.8 · UD-Q4_K_XL21.8 t/s
- Qwen3.8 · UD-Q3_K_XL23.5 t/s
- Qwen3.8 · UD-Q2_K_XL24.8 t/s
- Qwen3.8 · UD-IQ3_XXS24.2 t/s
- Qwen3.8 · UD-IQ2_XXS24.2 t/s
- Qwen3.6 · UD-Q4_K_M81.1 t/s
- Qwen3.5 · Q4_K_M68.0 t/s
- Qwen3.5 · Q4_K_M101.4 t/s
- Nemotron 3 Nano · Q4_K_M93.6 t/s
- Nemotron 3 Nano · Q4_K_M91.8 t/s
- Muse · UD-Q4_K_XL28.5 t/s
- Model not identified · Q4_K_M238.6 t/s
- gpt-oss · MXFP4123.1 t/s
- Gemma 4 · UD-Q4_K_M88.1 t/s
- gpt-oss · MXFP485.4 t/s
- Gemma 4 · Q4_054.5 t/s
- Falcon-H1 · Q4_K_M52.7 t/s
- Falcon-H1 · Q4_K_M86.4 t/s
- Falcon-H1 · Q4_K_M147.5 t/s
- Granite-4.0-H · Q4_K_M141.9 t/s
- Granite-4.0-H · Q4_K_M282.5 t/s
- Qwen3.8 · UD-Q4_K_XL12.0 t/s
- Qwen3.5 · Q4_K_M15.6 t/s
Frequently asked
How fast is llama.cpp on the M4 Max 128GB?
On the M4 Max 128GB, llama.cpp runs Qwen3.8 125B · 6B active at 3-bit at a median 27.3 tokens per second for generation, from 1 plain run (32K context).