Liquid AI has released LFM3, a small-model family with text, vision, audio, omni and decision checkpoints. The collection starts with a 1.73-billion-parameter text base and lists six post-trained checkpoints. The text checkpoint advertises a 256-thousand-token context window; the audio and omni cards list 32 thousand. Those are materially different operating envelopes within the same family.
Its license is immediately relevant to a large firm considering it. The LFM Open License does not license commercial use by a legal entity with more than ten million dollars in annual revenue, and the definition includes commonly controlled entities. Putting the project in a small subsidiary does not, on that wording, make the revenue condition disappear.
I’d be tempted to try it, once the license is resolved.
I’d leave you to enjoy the download.
The official safetensors listings are 3.46 gigabytes for text, 3.84 for vision-language, and 4.33 for omni. Those are file sizes; they don’t establish the memory footprint of serving a long context or handling the other modalities.
The model card discourages agentic coding and complex multistep reasoning, even as the release includes a decision checkpoint and reports WebArena results. Liquid’s edge-speed and benchmark claims come from named devices and settings and remain vendor-reported. For a deployment, memory, latency, task quality and security still need to be measured in the intended environment.