Liquid AI releases LFM3, a family of open 2B-parameter models
Six models for text, vision, audio, omni and decision tasks share one 1.7B-parameter backbone pre-trained on 22T tokens.
Source published · In the Daily Brief of 28 October 2026
Filed under Large language models
Liquid AI released LFM3 on 27 October, with open weights on Hugging Face. The backbone, LFM3-1.7B-Base, uses a hybrid of three operators across 24 layers: 13 gated short convolutions, five sliding-window attention layers with a 256-token window and six global attention layers. The company says it outperforms its earlier 2.6B-parameter model with 35% fewer parameters.
The text model has a 256K-token context window and fits in about 1 GB at 4-bit precision. Every model can be fine-tuned on a single GPU, and the architecture was chosen by a search that measured speed on 14 combinations of devices and back ends, from phones to cloud CPUs.