# llama.cpp update restores FP16‑to‑FP32 speed on Intel CPUs

> The Oct 15, 2026 release (b11635) keeps the FP16→FP32 lookup table on x86, fixing a slowdown in quantized kernels.

Oossa · 2026-10-15 · https://oossa.com/en/llama-cpp-update-restores-fp16-to-fp32-speed-on-intel-cpus

The open‑source llama.cpp library released version b11635 on 2026‑10‑15. It changes the CPU code to keep the FP16‑to‑FP32 lookup table on x86 processors that support the F16C instruction set. The prior change routed both scalar conversions through F16C, which slowed Q8_0/Q4_0 matrix‑multiply kernels by 45‑65 %. The new fix comments out the problematic define and restores the original table for FP16→FP32 conversions.

## The facts

- Release tag b11635 published on 2026‑10‑15
- Fix targets x86 CPUs with F16C, improving Q8_0/Q4_0 kernel speed by up to 65 %

## Why it matters

Developers running llama.cpp on Intel CPUs will get noticeably faster model inference after updating.

## Sources & references

1. [ggml-org/llama.cpp b11635](https://github.com/ggml-org/llama.cpp/releases/tag/b11635) – llama.cpp, 2026-10-15

Last updated: 2026-10-15