llamafile update adds local support for two very large models

Mozilla.ai has released llamafile v0.10.5, an update to its tool for packaging and running large language models locally. The release updates its llama.cpp core and adds support for two large models: the compressed 6GB Ternary Bonsai 27B and the 118B Laguna-S-2.1 coding mixture-of-experts model.

llamafile is aimed at making local model execution easier by bundling model-running software into portable files. That matters for developers and researchers who want to test open-weight models without sending prompts to a hosted service.

The update also fixes documentation and adds prebuilt transcribefile binaries for speech-to-text work. Those changes are smaller than the model support, but they reduce setup friction for people using the project across machines.

The practical limit remains hardware. Local execution can improve privacy and control, but very large models still need enough memory and compute to be useful at interactive speeds.