# llama.cpp adds SYCL weight reorder for faster IQ3 inference

> The open‑source llama.cpp library now reorders IQ3_S and IQ3_XXS weights at load time, boosting token throughput on Intel Arc GPUs.

Oossa · 2026-10-15 · https://oossa.com/en/llama-cpp-adds-sycl-weight-reorder-for-faster-iq3-inference

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.

## The facts

- Release date: 2026-10-15 ("Thu Oct 15 2026 06:50:37 GMT+0200")
- Token speed on Intel Arc Pro B70 improved from 12.92 t/s to 23.54 t/s (+82 %) for IQ3_S

## Why it matters

Developers running llama.cpp on SYCL‑compatible GPUs can get noticeably faster responses without changing model code.

## Sources & references

1. [ggml-org/llama.cpp b11636](https://github.com/ggml-org/llama.cpp/releases/tag/b11636) – llama.cpp, 2026-10-15

Last updated: 2026-10-15