llama.cpp b9660 fixes tool-call double-escaping in local LLM parsing

The popular open-source C++ project llama.cpp, designed to run large language models locally, has shipped a quick patch release (b9660) addressing a double-escaping issue in its tool-call parsing for LFM2 models. The bug caused escape characters in function arguments to be incorrectly doubled (e.g., \" becoming \\"), leading to malformed JSON and failed tool invocations in agent systems. The release also adds new test cases to prevent regression.

This fix is particularly important for developers building autonomous agents or RAG pipelines that rely on local LLMs for tool use. llama.cpp supports a wide range of hardware, including Apple Silicon, Intel, ARM64, and GPU acceleration via Vulkan, CUDA 12/13, ROCm, SYCL, and OpenVINO. The project, maintained by ggml-org, remains one of the most popular ways to run quantized LLMs offline, with 117,000 GitHub stars. The patch is immediately available as a prebuilt binary for all major platforms.