# llama.cpp adds BF16 support for RMS_NORM on CPU and CUDA

> The open‑source Llama inference library now handles BF16 precision in its RMS_NORM operation, with binaries for macOS, iOS and Linux.

Oossa · 2026-11-13 · https://oossa.com/en/llama-cpp-adds-bf16-support-for-rms-norm-on-cpu-and-cuda

The ggml‑org team released llama.cpp version b12366 on 2026‑11‑13. The update adds BF16 (a 16‑bit floating‑point format) support to the RMS_NORM layer for both CPU and CUDA back‑ends. A single helper now launches the F32/BF16 RMS_NORM kernels. Pre‑built binaries for macOS Apple Silicon, Intel, iOS, and several Ubuntu variants are available for download.

## The facts

- Release version b12366 published 2026‑11‑13
- BF16 support added to RMS_NORM for CPU and CUDA

## Why it matters

Developers can run Llama models faster on hardware that prefers BF16, lowering memory use without losing much accuracy.

## Sources & references

1. [ggml-org/llama.cpp b12366](https://github.com/ggml-org/llama.cpp/releases/tag/b12366) – llama.cpp, 2026-11-13

Last updated: 2026-11-13