Liquid AI releases the LFM3 model family

A family for local adaptation

Liquid AI released six roughly two-billion-parameter post-trained models on a shared backbone, covering text, vision, audio, omni and two decision variants. The launch supplied downloadable checkpoints; Webdiga reported same-day local text and audio tests while distinguishing runtime support for each modality. [1][2]

The weights use the LFM Open License v1.0. Its commercial conditions apply at annual revenue of $10 million or more, with separate provisions for research and qualified nonprofits. Download access does not establish unrestricted commercial licensing, and the text-model card cautions against agentic coding and complex multi-step reasoning. [3][4]