Liquid AI releases LFM3, six open-weight 2B-class models for text, vision, audio and decisions

Liquid AI has published the third generation of its Liquid Foundation Models. LFM3 arrived on October 27, 2026 as a family of six small open-weight models that all derive from the same 1.7-billion-parameter base.

What was released

The lineup covers most modalities in one go. LFM3-1.7B is the text model, with a 256,000-token context window. LFM3-VL-2B handles images and, for the first time in the family, video. LFM3-Audio-2B takes speech in and produces speech out, in 18 languages according to the company. LFM3-Omni-2B combines vision, audio and language. Two "decision" models, d1-2B and d1-Omni-2B, are tuned versions of the vision and omni models designed to return calibrated answers instead of free text.

Base and post-trained weights are on Hugging Face, and the models can also be used through Liquid's Playground and API. The model card for the text model lists the LFM Open License v1.0 and support for 16 languages. Liquid names AMD, Arm, Intel, NVIDIA, Qualcomm, OpenRouter and Vercel among its launch partners, and says Mercedes-Benz and Shopify tested pre-release versions.

What the company claims

The pitch is efficiency. Liquid says the text model needs about 1 GB of memory at 4-bit precision and that a full 128K-token context requires a 1.6 GB cache, well below the figures it reports for comparable Qwen and MiniCPM models. It also claims the 1.7B model outperforms its previous 2.6B model and posts the best average score in its comparison table, ahead of the much larger Qwen3.5-9B, with particularly strong results on tool-calling and web-agent tests. For vision, it says LFM3-VL-2B lands close to Gemma 4 12B across 62 benchmarks.

All of these numbers come from Liquid AI and have not been independently reproduced.

Limits to keep in mind

The company's own blog acknowledges that the text model trails larger models on coding and knowledge tasks, and the model card says it is not recommended for agentic coding or complex multi-step reasoning, suggesting retrieval for knowledge-heavy work. The release fits a broader push toward compact models that run locally on phones, cars and laptops, where memory and latency matter more than raw capability.