b9062

The llama.cpp project, known for its efficient local execution of large language models, has pushed release b9062. The key change is a fix in the common/chat module to ‘preserve media markers for typed-content templates’ (PR #22634). This means that when users interact with models via chat templates that include typed content (e.g., images or video placeholders), those markers are no longer stripped or corrupted across turns. The release includes pre-built binaries for a wide range of hardware: Apple Silicon with and without KleidiAI, Intel macOS, iOS XCFramework, multiple Linux variants (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, SYCL, HIP), and Android arm64.

With over 109,000 GitHub stars and 17,900 forks, llama.cpp is one of the most critical open-source tools for running AI models offline. This incremental update highlights the community’s attention to multimodal support—a growing requirement as LLMs increasingly handle images, audio, and video. By preserving media markers, developers can build richer chat experiences without workarounds, moving closer to plug‑and‑play local AI assistants that understand context beyond text.