b8799

The open-source project llama.cpp, maintained by ggml-org, has released a new commit (b8799) that significantly expands its model compatibility and hardware support. The key technical update is an enhancement to the 'autoparser' system, which now supports the 'JSON_NATIVE' format with per-call markers. This change was specifically implemented to enable full compatibility with the Reka-Edge model from the AI startup Reka AI, allowing users to efficiently run this new, capable model on their local machines using the optimized llama.cpp engine.

Alongside this model support, the release includes a major expansion of its pre-built binary distribution matrix. Developers can now download builds for Windows x64 with updated CUDA 13.1 DLLs, adding to the existing CUDA 12 support. The release also adds binaries for Vulkan backends on both Ubuntu x64/arm64, an OpenVINO build for Ubuntu x64, and maintains its comprehensive support for macOS Apple Silicon, iOS, and various Linux flavors. This broadens the tool's accessibility for production deployment across different environments.

This update underscores llama.cpp's role as a critical infrastructure layer in the open-source AI ecosystem. By quickly integrating support for new models like Reka-Edge and staying current with the latest GPU compute libraries, it lowers the barrier for developers and researchers to experiment with state-of-the-art models without relying on proprietary APIs. The project's commit-based release model allows for rapid iteration and community testing of new features.