llama.cpp SYCL speed results for local models
llama.cpp build dd4623a74(9640) was measured with several Gemma-family models running on the SYCL backend. Q8_0 is 4.69GiB with 4.65 billion , and reached 5662.45 on pp512 and 109.14 on tg128. gemma4 12B Q8_0 is 11.78GiB with 11.91 billion , and reached 1578.19 on pp512 and 32.43 on tg128.
gemma4 26B.A4B Q8_0 is 25.00GiB with 25.23 billion , and reached 1332.35 on pp512 and 40.13 on tg128. A qwen35moe 35B.A3B Q8_0 entry is also listed at 34.36GiB and 34.66 billion , but the provided text cuts off before its final speed result.
Key points
- Q8_0 was the fastest shown for generation, at 109.14 on tg128.
- gemma4 12B Q8_0 reached 32.43 on tg128.
- gemma4 26B.A4B Q8_0 reached 40.13 on tg128.
- Smaller may be better for low-cost AI agent experiments when speed matters.
- The benchmark does not show answer quality or real agent task success.