The llama.cpp project released version b11636 on 2026-10-15. It adds a persistent weight layout for the IQ3_S and IQ3_XXS quantizations when running under SYCL. The weights are reordered once when the model loads, letting the GPU process dense blocks instead of gathering bits, which speeds up decoding and prefill. On an Intel Arc Pro B70 GPU the Qwen3.8‑27B model saw token rates rise from 12.92 to 23.54 tokens per second for short prompts, an 82 % increase.
Why it matters
Developers running llama.cpp on SYCL‑compatible GPUs can get noticeably faster responses without changing model code.
