Liquid AI's LFM3 reads text, video and audio on a phone

Liquid AI published LFM3 on October 27, 2026, a family of six models near 2 billion parameters that share one backbone and cover text, images, audio, video and decision tasks. The company announced the family in a blog post, with a companion write-up on Hugging Face and the open weights in a Hugging Face collection.

The six models all branch from one base, LFM3-1.7B-Base, pre-trained on 22 trillion tokens: a text model, a vision-language model, an audio model, an omni model, and two "decision" models, d1-2B and d1-Omni-2B, that return typed answers instead of prose. Liquid AI is a lab whose Liquid Foundation Models are built to run on phones, laptops and cars rather than in a data center.

The video claim needs the fine print. In LFM3 the video capability is input, not generation. LFM3-VL-2B is the first Liquid vision model to take video as input, its model card states, and it answers in text. The card says video is sampled at 2 frames per second with 4 to 768 frames by default. The omni model reads video alongside its soundtrack, sampled at 2 fps with 4 to 384 frames. Nothing in the release generates video.

Liquid AI says the text model fits in about 1 GB at 4-bit precision with a 256K-token context, that every model fine-tunes on a single GPU, and that LFM3-1.7B beats the older LFM2.5-2.6B while using 35 percent fewer parameters. The company reports its vision model lifted an average across 62 benchmarks from 57.6 to 71.4 over LFM2.5-VL-3B.

None of those figures has been independently verified. The explainer site eesel.ai notes that Artificial Analysis had not listed LFM3 as of October 28, and that every number is Liquid's own. The vision model's card lists vLLM support as in progress and SGLang support as in review. A separate dev.to write-up that did not run the models says the same about the benchmarks.

The weights are open under Liquid's LFM Open License v1.0, which liquid.ai says is free, including for commercial use, until a company passes $10 million in annual revenue. VentureBeat reports API pricing for the text model of 8 cents per million input tokens and 20 cents per million output, and a 128,000-token hosted context against the model card's 256K.

For a creator, a small model that reads footage locally points at cheap private first passes: a rough read of a clip or a scratch voiceover that never leaves the machine. Whether that video understanding holds up outside Liquid AI's own charts is still untested.

Download the LFM3 weights from the Hugging Face collection, try them in the Liquid Playground, or run the text model through llama.cpp or MLX.