llama.cpp Release b11486 Improves GELU Accuracy on Qualcomm Hexagon Backend
llama.cpp version b11486 was released, featuring a pull request update (#30104) that improves the numerical accuracy of the GELU activation function when executing on Qualcomm Hexagon backends. The improvement was generated with assistance from OpenCode. Improving mathematical accuracy for core activation functions on specialized edge hardware prevents quality degradation when running LLMs locally on Snapdragon-powered mobile and embedded devices. As edge AI adoption grows, maintaining precise low-level math ops across non-x86 backends like Hexagon DSPs is vital for model performance stability. The release provides compiled binaries for various target environments, including Linux arm64 and Android builds designed specifically for Qualcomm Snapdragon hardware combining CPU, Adreno GPU, and Hexagon NPU support. Additionally, KleidiAI optimizations remain explicitly disabled for the macOS Apple Silicon arm64 release binaries.
## BACKGROUND
GELU (Gaussian Error Linear Unit) is a widely adopted activation function used in modern Transformer-based language models to introduce non-linearity. Qualcomm Hexagon is a digital signal processor (DSP) and neural processing unit (NPU) architecture built into Snapdragon processors for power-efficient AI compute. llama.cpp is a popular open-source inference library designed to run large language models locally across diverse hardware architectures.