llama.cpp b9625 fixes Jinja templating bug in negative step slicing

llama.cpp, the popular open-source C++ library for running large language models locally, has released version b9625. This patch primarily addresses a bug in Jinja template parsing where negative step slices combined with start/stop values produced incorrect behavior. Jinja templates are commonly used to format prompts for chat models, so this fix is critical for developers relying on precise prompt construction.

The release includes pre-built binaries for a wide range of platforms: Apple Silicon (ARM64) with and without KleidiAI acceleration, Intel Macs, Linux (x64, ARM64, s390x), Android ARM64, and Windows (x64, ARM64). GPU acceleration support spans CUDA (12 and 13), Vulkan, ROCm, SYCL, and OpenVINO, making it one of the most portable options for local AI inference. The project now boasts 116k stars and 19.6k forks on GitHub, cementing its role as a cornerstone of the local LLM ecosystem.

With b9625, developers can confidently update their local setups to avoid template-related errors in multi-turn conversations or structured outputs. This incremental improvement reflects the broader trend of maturing local AI tooling, where stability and broad hardware support are as important as raw performance gains.