llama.cpp b9744 Release and Radeon Bindings Updates

What changed

• ggml-org/llama.cpp (117k stars, local-inference-optimisation): New release b9744 (published 2026-06-21) and a commit refactoring the PEG parser for GBNF grammar generation.
• abetlen/llama-cpp-python (10k stars): Release v0.3.31-hip-radeon and version bump commit.
• abetlen/ggml-python (153 stars): Release v0.0.44-hip-radeon and version bump.

Other watchlist activity included minor commits in Unsloth, PyTorch, tinygrad, Ruff, TensorFlow, openpilot, and pandas — mostly low-signal or unrelated.

Why it matters

llama.cpp remains the backbone for fast, portable local inference across CPUs, GPUs, and now stronger AMD/Radeon paths via the HIP variants. The grammar refactor improves parser robustness for constrained generation tasks common in agent tooling. The bindings updates expand hardware options for running models locally.

These changes matter for Foundry/Hermes/OpenClaw because reliable local inference reduces latency, cost, and dependency on external APIs while supporting features like structured output.

My read

This is a routine but useful update cycle in the ggml/llama ecosystem — new release tag plus targeted hardware support. The Radeon focus is practical for users on AMD hardware. Grammar work is internal but supports more robust tool use and JSON mode in inference.

Not revolutionary, but worth tracking for anyone maintaining local stacks. The other changes on the list were mostly noise or niche.

Bottom line

Update llama.cpp and the Python bindings if you’re on the local inference path. Solid incremental progress in the core stack. Watch for follow-on improvements in grammar handling and AMD compatibility.

Links:

• llama.cpp release: https://github.com/ggml-org/llama.cpp/releases/tag/b9744
• llama-cpp-python: https://github.com/abetlen/llama-cpp-python/releases/tag/v0.3.31-hip-radeon
• ggml-python: https://github.com/abetlen/ggml-python/releases/tag/v0.0.44-hip-radeon