# llama.cpp b12122 adds a fused CUDA operation for NVFP4

> The release combines two model-computation steps in one CUDA path for NVFP4, a low-precision number format. Builds are available for several platforms.

Oossa · 2026-11-05 · https://oossa.com/en/llama-cpp-b12122-adds-a-fused-cuda-operation-for-nvfp4

The llama.cpp project has released version b12122, which adds a CUDA optimization that fuses NVFP4 gate/up matrix multiplications with a GLU operation in MMQ, its matrix-multiplication path. NVFP4 is a compact number format used in some AI computations; combining steps can reduce separate work, though the release notes give no performance measurements.

The release also includes downloadable builds for macOS, iOS and Ubuntu on x64, Arm64 and s390x.

## The facts

- Release: llama.cpp b12122; the listed CUDA change is pull request #31296.
- Downloads include Ubuntu x64, Arm64 and s390x CPU builds, plus macOS and iOS packages.

## Why it matters

People running supported NVFP4 workloads on CUDA may get the fused computation path, but the release notes do not quantify any speed or memory gains.

## Sources & references

1. [ggml-org/llama.cpp b12122](https://github.com/ggml-org/llama.cpp/releases/tag/b12122) – llama.cpp, 2026-11-05

Last updated: 2026-11-05