# llama.cpp b12094 speeds up Hexagon top-k sampling

> The llama.cpp release changes top-k sampling on Hexagon hardware, using a scan to avoid sorting blocks that cannot make the cut. No speedup figures are given.

Oossa · 2026-11-04 · https://oossa.com/en/llama-cpp-b12094-speeds-up-hexagon-top-k-sampling

The ggml-org team released llama.cpp b12094 on November 4, 2026, with a faster top-k sampling path for its Hexagon backend. Top-k sampling selects among the highest-scoring options when generating text.

The change targets k values from 2 to 64 and rows wider than 1,024 entries. It skips sorting blocks that cannot contain a top result, but the release gives no measured speedup.

## The facts

- The new path uses a Hexagon vector max operation to screen blocks of 256 values.
- It applies to both single-row and multi-row processing.

## Why it matters

People running llama.cpp on supported Hexagon hardware get a more efficient sampling method for these workloads, though the release does not quantify any real-world speed gain.

## Sources & references

1. [ggml-org/llama.cpp b12094](https://github.com/ggml-org/llama.cpp/releases/tag/b12094) – llama.cpp, 2026-11-04

Last updated: 2026-11-04