# llama.cpp expands BF16 execution across CPU and CUDA

This edition covers a wider window since the previous 16:00 UTC publication. llama.cpp added BF16 variants for core element-wise operations on CPU and CUDA, groundwork for keeping model activations in BF16 during inference.

## llama.cpp adds BF16 unary, GLU, binary and scale operations

llama.cpp release b11293 adds BF16 variants of unary, GLU, binary and scale operations across its CPU and CUDA backends. The change is intended to let later inference paths keep activations in BF16 rather than converting them around unsupported element-wise operations. The merged implementation includes BF16 support for operations including activations, gated linear units, binary arithmetic and scaling, extending the runtime's native BF16 execution coverage and laying groundwork for broader BF16 activation support. 

[llama.cpp b11293 release](https://github.com/ggml-org/llama.cpp/releases/tag/b11293) · [llama.cpp PR #29675 — add BF16 unary, GLU, binary and scale ops](https://github.com/ggml-org/llama.cpp/pull/29675)
