llama.cpp on the Intel Arc Pro B70
On the Intel Arc Pro B70, llama.cpp runs Qwen3.6 35B · 3B active at 4-bit at a median 63.7 tokens per second for generation, from 5 plain runs (127K to 256K context, 2 not stated).
Speed by model on the Intel Arc Pro B70
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 4-bit | 63.7 t/smedian of 5, 127K to 256K context, 2 not stated | no other engine reported |
| Qwen3 8B 4-bit | 77.7 t/sone report, 2K context | no other engine reported |
| Qwen3.8 27B 4-bit | 25.6 t/sone report, context not stated | no other engine reported |
| Qwen3.8 27B 8-bit | 16.7 t/sone report, context not stated | no other engine reported |
11 llama.cpp reports on the Intel Arc Pro B70
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.
- Qwen3.6 35B · 3B active · Q4_K_M20.7 t/s
- Qwen3.8 27B · UD-Q8_K_L16.7 t/s
- Qwen3.8 125B · 6B active · Q2_031.2 t/s(speculative decoding, part in system RAM, 3 requests at once)
- Qwen3 8B · Q4_K_M77.7 t/s
- Qwen3.6 35B · 3B active · UD-Q4_K_M54.7 t/s
- Qwen3.6 35B · 3B active · Q4_K_M76.0 t/s
- Qwen3.8 27B · Q4_K_M25.6 t/s
- Qwen3.8 27B · Q4_K_M35.6 t/s(speculative decoding)
- Qwen3.6 35B · 3B active · Q4_K_XL63.7 t/s
- Qwen3.6 35B · 3B active · Q4_K_M70.5 t/s
- Qwen2.563.0 t/s
Frequently asked
How fast is llama.cpp on the Intel Arc Pro B70?
On the Intel Arc Pro B70, llama.cpp runs Qwen3.6 35B · 3B active at 4-bit at a median 63.7 tokens per second for generation, from 5 plain runs (127K to 256K context, 2 not stated).