Mozilla AI released llamafile 0.10.5, bringing its single-file local-model runtime up to date with a fast-moving llama.cpp base and several model shapes that stress ordinary inference packaging.
The release calls out Ternary Bonsai 27B, whose weights use the values -1, 0, and +1 and fit in roughly 6 GB, including a vision component that can be loaded only when needed. It also adds support for Poolside's Laguna-S-2.1, a 118B mixture-of-experts model that activates about 8B parameters per token; the quantized checkpoints expose 256K context rather than the BF16 model's 1M context.
The less visible signal is release engineering. Mozilla synchronized three upstream changes in two weeks and improved an agent skill that drafts the synchronization pull requests. A prebuilt transcribefile speech-to-text binary now ships as a release artifact as well.
This is not a new inference architecture by itself. It is evidence that a useful local-AI product is a compatibility and distribution operation: new kernels, formats, modalities, and executable packaging have to arrive as one reproducible runtime.