# llama.cpp release b11688 adds CPU AVX2 speed tweak

> The open‑source llama.cpp library ships a new binary with reduced AVX2 tiled accumulators, improving CPU performance.

Oossa · 2026-10-16 · https://oossa.com/en/llama-cpp-release-b11688-adds-cpu-avx2-speed-tweak

The ggml‑org team released llama.cpp version b11688 on 2026-10-16. The update adds a CPU‑only change that reduces AVX2 tiled accumulators, a tweak the project says speeds up inference. Pre‑built binaries for macOS, iOS, Linux, Vulkan and CUDA are provided for a range of hardware. The change is labeled as assisted by Codex, the code‑generation model from OpenAI.

## The facts

- Release date: 2026‑10‑16 ("Fri Oct 16 2026 12:03:55 GMT+0200")
- New CPU optimization: reduce AVX2 tiled accumulators

## Why it matters

Developers can run llama.cpp models faster on AVX2 CPUs without rebuilding the code.

## Sources & references

1. [ggml-org/llama.cpp b11688](https://github.com/ggml-org/llama.cpp/releases/tag/b11688) – llama.cpp, 2026-10-16

Last updated: 2026-10-16