llamafile v0.10.5

llamafile v0.10.5 (released August 5, 2026) now supports two large models that can run locally on a laptop or desktop thanks to an updated embedded llama.cpp core. Ternary Bonsai 27B, a compressed model from PrismML using ternary weights, requires only about 6 GB on disk while retaining most of the quality of the full-precision Qwen3.6-27B base. Laguna-S-2.1, Poolside’s 118B-parameter coding Mixture-of-Experts, activates only 8B parameters per token, so a quantized version fits in consumer memory and runs interactively at usable speeds. Both models were previously available as GGUF weights but could not be loaded by older llamafile builds because the bundled llama.cpp lacked support for their architectures. The release also fixes documentation on the help system, release binaries, and GPU support, and adds pre-built transcribefile speech‑to‑text binaries as a community contribution. The underlying sync work—three upstream merge passes in two weeks, with an improved automation agent—should accelerate support for new models in future releases.