LFM3

Liquid AI's open-weight 2B-class models for on-device text, vision, audio, and local agents.

Also known as LFM3-1.7B, Liquid Foundation Models

What it does

LFM3 is Liquid AI's family of open-weight models, announced October 27, 2026. The post says the first release covers six models around 2 billion parameters, branched from LFM3-1.7B-Base, for text, vision, audio, omni, and decision work. LFM3-1.7B is the text model for local agents. The model card lists 1.73 billion parameters, a 256K-token context window, 24 layers, and a 128,000-token vocabulary, and says it is not recommended for agentic coding because of its size. The announcement says the text model fits in about 1 GB at 4-bit. Decision models named in the post are d1-2B, from LFM3-VL-2B, and d1-Omni-2B, from LFM3-Omni-2B. Weights use the LFM Open License v1.0, which allows commercial use only when annual revenue stays under $10 million. Liquid says the models are on Hugging Face, its Playground, and the Liquid AI API. The post does not list an API price.

Who is LFM3 for?

People running a small on-device model for local agents, vision, audio, or a narrow fine-tune.

Tags

• open-weight
• local
• agents