Liquid AI's LFM3 splits into six small models, one per modality
The new family covers text, vision, audio and omni in six roughly 2B-parameter models, with a 256K-token context window and open weights on Hugging Face.
Liquid AI released LFM3 on 27 October, a family of six open-weight models built on a new “tri-hybrid” architecture that mixes gated short convolutions, sliding-window attention and global attention across 24 layers. Each model sits around 2 billion parameters, branched from a shared LFM3-1.7B-Base that was pre-trained on 22 trillion tokens, and the lineup splits by modality: a text model, a vision model that takes images and video, an audio model, an omni model that takes all of the above, and two “decision” variants that output a calibrated choice rather than free text.
The text model, LFM3-1.7B, carries a 256K-token context window and outperforms the larger LFM2.5-2.6B while using 35% fewer parameters. All six ship as open weights on Hugging Face under Liquid’s own LFM Open License v1.0, the text model runs in roughly 1GB at 4-bit precision, and Liquid says it is built to run on anything from a phone to a cloud CPU.
LFM3-1.7B text
LFM3-VL-2B text, image, video -> text
LFM3-Audio-2B text, audio -> text, speech
LFM3-Omni-2B text, image, video, audio -> text, speech
d1-2B text, image, video -> decision
d1-omni-2B text, image, video, audio -> decisionWhy it matters: this is Liquid’s third model release this month, and the pattern is consistent: small, specialised models meant to run on-device rather than through an API, with a context window and modality mix that used to need a much bigger model. If you’re building anything that has to run locally, on a laptop, phone or edge box, this is a real set of options rather than a scaled-down demo.
The caveat: the LFM Open License v1.0 is not Apache or MIT. Liquid’s earlier LFM2.5 release under the same license allowed free commercial use only below $10 million in annual company revenue, so check the licence text before shipping this in a product.