Build specialized multimodal models with LFM3
Liquid AI introduces LFM3, a family of roughly 2-billion-parameter models for text, vision, audio, omni and decision tasks on one shared backbone. The text model reportedly beats its predecessor LFM2.5-2.6B with about a third fewer parameters; weights are open, for single-GPU fine-tuning and on-device use.
What the article says
On October 27, 2026, Liquid AI presented LFM3, the third generation of its Liquid Foundation Models, on the Hugging Face blog. The family covers text, vision, audio, omni and decision models, with six models of about 2 billion parameters built on one shared backbone; LFM3-1.7B, at 1.7 billion parameters, outperforms LFM2.5-2.6B with 35% fewer parameters.
The text model fits in about 1 GB at 4-bit precision and has a 256K-token context window, aimed at local agents and tool calling. LFM3-VL-2B accepts video for the first time, LFM3-Audio-2B handles voice across 18 languages, LFM3-Omni-2B is the company's first omni model, and the d1 models answer several questions at once with calibrated probabilities. Every model can be fine-tuned on a single GPU, Colab notebooks are provided, and all weights are open on Hugging Face.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.