The details
All six models branch from one base model, LFM3-1.7B, trained on 22 trillion tokens. A new architecture mixes short convolutions with two kinds of attention, which Liquid AI says makes it about 25% faster with 8K tokens of context. On the WebArena-Verified agent test, the text model scored 38.42, against 3.02 for its predecessor. The vision model averages 71.4 across 62 benchmarks, up from 57.6. Liquid AI also adapted the vision model into robotics models that generate actions in under 50 milliseconds on a Jetson board.
Why it matters
Small open models that a company can fine-tune cheaply and run on its own devices keep data local and avoid cloud serving costs. One shared backbone also makes it simpler to add voice or vision to an app. The benchmark results are Liquid AI’s own, so independent tests will show how they compare.