llama.cpp b10375 tightens Qwen function parsing for reliable AI tool calls

ggml-org's llama.cpp, the de facto standard for running LLMs locally on consumer hardware, received a new release tagged b10375 on August 12. The update is modest but meaningful: it tightens bare function parsing for Qwen models via pull request #26793. In practice, this addresses a longstanding quirk where Qwen models sometimes emit function calls without all required arguments or formatting, causing errors in agentic workflows. The fix ensures that chat completions involving function/tool calling are parsed correctly, so applications can reliably execute the model's intended actions.

The release ships with prebuilt binaries for an extensive range of platforms and backends: macOS with Apple Silicon (including a KleidiAI-optimized build), Intel x64, iOS, Linux on x64/arm64/s390x, Android ARM64, and Windows x64/arm64. GPU acceleration options include Vulkan, CUDA 12.4/13.3 DLLs, ROCm 7.14, OpenVINO, SYCL (FP16/FP32), and Adreno OpenCL. The project now boasts over 124,000 stars and 21,600 forks, reflecting its status as the go-to engine for on-device inference. For developers running Qwen models like Qwen2.5 or Qwen3 for tool use, upgrading to b10375 is a low-risk way to improve stability in production agent pipelines.