Coding Tools

llama.cpp b11272 optimizes Hexagon NPU and adds CUDA 13

llama.cpp ReleasesSeptember 30, 2026high confidence

Why it matters

→Hexagon NPU optimizations improve efficiency for AI inference on Snapdragon devices.
→CUDA 13 support ensures compatibility with the latest NVIDIA driver stacks.
→ROCm 10.0 inclusion maintains viable AMD GPU support for local inference.

llama.cpp has released build b11272, focusing on performance optimizations for Qualcomm Hexagon NPUs and expanded GPU driver support. Key changes include optimizing the concat operation by reducing packet overhead in gather/transpose hot loops and replacing software division with fast paths. The release adds official support for CUDA 13 (13.4 libraries) on Linux and Windows, alongside continued ROCm 10.0 support for AMD GPUs. Notably, macOS Apple Silicon builds with KleidiAI are currently disabled in this version.

Read original

llama.cpp b11272 optimizes Hexagon NPU and adds CUDA 13

Why it matters

More from llama.cpp Releases

llama.cpp fixes MoE dispatch inefficiency on Vulkan

llama.cpp Vulkan perf boost on Intel Arc

More in Coding Tools

Claude Code v2.1.284 adds Sonnet 5.5

Claude Code v2.1.285 patches security and stability

llama.cpp b11267 adds ROCm 10 and Snapdragon support

GitHub adds repo-level Dependabot runner config