Liquid AI introduces LFM3: small models for text, vision and speech

Liquid AI introduced its LFM3 family on October 27, 2026, the same day this article was published. The company describes a group of roughly two-billion-parameter models for text, vision, audio and combined modalities. It says open-weight models are available through its Hugging Face collection and hosted services.

What the family does

The company’s technical overview lists LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B for audio and text, and LFM3-Omni-2B for mixed inputs and speech. Separate decision models use the same backbone. The different checkpoints have different inputs, outputs and runtime requirements: picking a text model does not automatically give it speech or vision.

Why local users may care

Liquid AI says the text model’s weights take about 1 GB at four-bit precision and its context window is 256K tokens. That weight estimate is not total memory use: context, runtime, operating system and other components can add substantially to requirements. Large context also does not ensure fast or accurate performance. We have not tested the family, its hardware fit or its speed in the AiBrainWorX launcher.

The LFM3-1.7B model card describes it as suitable for local agents and structured output, while explicitly advising against agentic coding and complex multi-step reasoning for this small model. It lists the LFM Open License v1.0; open weights do not automatically mean MIT or unrestricted redistribution. Check each checkpoint’s own license and runtime instructions before use. Some examples request trust_remote_code=True, which executes downloaded model code, so review the source and choose an appropriate environment.