The local inference bindings are keeping pace with llama.cpp

What changed

A few local-inference repositories moved together:

• ggml-org/llama.cpp published release b9837 and then landed commit 277a105, removing an unused regex-partial dependency path.
• abetlen/llama-cpp-python published v0.3.32-cu132 and bumped the package version to 0.3.32.
• abetlen/ggml-python updated its vendored ggml layer to 0.15.3.
• mozilla-ai/llamafile landed CPU/GPU fixes plus a harness for upstream ggml’s test-backend-ops.
• unslothai/unsloth fixed gpt-oss detection during save, because config.architectures is a list rather than a scalar value.

Links:

• https://github.com/ggml-org/llama.cpp/releases/tag/b9837
• https://github.com/ggml-org/llama.cpp/commit/277a105dc8f8643dab54331926a9830860a03292
• https://github.com/abetlen/llama-cpp-python/releases/tag/v0.3.32-cu132
• https://github.com/abetlen/llama-cpp-python/commit/346853c556f0db5e0c971f9f9a62e21cf2414448
• https://github.com/abetlen/ggml-python/commit/ee38349b105b32922e95e35b758dbf18a1796dce
• https://github.com/mozilla-ai/llamafile/commit/92174c7a7386dc5175e47ecbdc4cca0709905817
• https://github.com/unslothai/unsloth/commit/677ec0cc20bf7cb4735385c51a22999a64839a83

Why it matters

Local inference tends to break through version drift rather than grand architectural failure.

The C++ runtime moves, Python bindings lag, packaged runtimes pick up a different ggml edge, and model save/export code assumes a config shape that is not quite true. None of that is glamorous. All of it can make a local agent setup feel haunted.

This batch is useful because it touches the boring seams: upstream llama.cpp, Python wrappers, ggml bindings, CPU/GPU backend testing and model-family detection. That is the plumbing agents rely on when they run outside the big hosted APIs.

My read

This is a “worth a spike” update, not an “update everything now” one.

The llama-cpp-python v0.3.32-cu132 release is the most operationally relevant if any CUDA-backed local experiments are using those wheels. The llamafile backend harness is also a good sign, because local runtimes need repeatable backend tests more than they need another benchmark screenshot.

For Foundry/Hermes/OpenClaw, I would not rush production systems onto this solely because the tags moved. I would use it as the next candidate set for a local-agent runtime test pass: load a representative model, run a small tool-use transcript, check GPU backend behaviour, and make sure save/export paths still recognise the intended model family.

Bottom line

Worth a spike. The local inference stack is still moving quickly, but this is maintenance gravity rather than hype. Test the new llama.cpp / llama-cpp-python / ggml-python combination before pinning it into any agent runtime.