Llama.cpp b10373 ships with MoE quantization fix and new platform builds

llama.cpp, the lightweight C++ library for running large language models on consumer hardware, has dropped a new release: b10373. Maintained by ggml-org, the project crossed 124k stars and 21.6k forks, and this latest tag packages a targeted optimization for quantization workflows. The headline commit, "imatrix.cpp: Move finite check and only check touched experts," touches the importance matrix (imatrix) code used during model quantization. Specifically, it moves a finite-value validation earlier in the pipeline and restricts checks to only the experts actually activated in Mixture-of-Experts (MoE) models. This can reduce computational overhead and improve quantization quality for sparse models like Mixtral or Qwen-MoE.

The release itself is a full distribution, offering pre-built binaries for a striking range of environments. On macOS, users get Apple Silicon (arm64) builds plus a KleidiAI-optimized variant, while Intel x64 and iOS XCFramework are also covered. Linux users can grab Ubuntu packages for CPU, Vulkan, ROCm 7.14, OpenVINO, and SYCL (FP32/FP16). Windows builds span CPU, CUDA 12 and 13 (with matching DLLs), Vulkan, OpenVINO, SYCL, and ROCm 7.14. Even Android (arm64 CPU) and openEuler (x86/aarch64 with Ascend ACL Graph) get dedicated packages. For developers, this means they can pull a compiled llama.cpp for nearly any edge device or server setup without compiling from source, while the imatrix refinement hints at fewer quantization artifacts for MoE models.