Liquid AI releases six small multimodal models built to run on devices
The open-weight LFM3 family covers text, vision, audio, omni input and decision tasks with models around two billion parameters.
Liquid AI has released LFM3, a family of six open-weight models designed for specialized AI that can run on phones, laptops and edge hardware. The lineup includes a 1.7-billion-parameter text model, a vision-language model, an audio model, an omni model combining vision, audio and language, and two decision models that return typed outputs with calibrated probabilities.
All six models are around two billion parameters and share a new backbone trained on 22 trillion tokens. The text model supports up to 256,000 tokens of context. Liquid says the models can be fine-tuned on a single GPU and has optimized the architecture across 14 hardware combinations, including mobile chips, CPUs, GPUs and NPUs.
LFM3 mixes gated short convolutions, sliding-window attention and global attention. The goal is to retain enough global context for agent tasks while keeping inference efficient on constrained devices. Liquid reports that its 1.7B text model is about 25% faster than the prior 1.2B model with an 8,000-token context, while its vision model accepts video.
Developers can deploy one compact model for a specific modality instead of sending every request to a large cloud model. The base and post-trained weights are on Hugging Face, with a hosted API, playground and local deployment documentation.
