Ollama v0.40.12 Released with Support for Liquid AI LFM3 and d1:2b

Ollama v0.40.12 adds support for Liquid AI's new LFM3 model family. Built on Liquid's new tri-hybrid architecture, LFM3 is designed for local agents, tool calling, and structured output. The family includes lfm3, a 1.7B text model with a 256K-token context window; lfm3-vl, which adds image understanding; lfm3-audio, providing speech recognition, translation, and text-to-speech in 18 languages; and lfm3-omni, which combines text, images, and audio in one model. Additionally, the 2B size decision models d1, built on LFM3, and d1-omni, which adds audio, are now supported.

In this release, lfm3 and d1:2b are supported on Apple hardware, with support for Linux/Windows coming soon. Other changes include faster prompt processing for Gemma 4, Flint, and Laguna models on Apple silicon; speculative decoding support for DFlash2 draft models on MLX; improved detail retention in small images for Gemma 4; ollama launch working with OpenCode v2; fixed tool-only assistant turns from Kimi CLI; and updates to MLX and llama.cpp.