Home/New releases/LFM3

LFM3

Liquid AI's open-weight 2B-class model family (text, vision, audio, omni, decision) built to run and fine-tune on-device.

Liquid AI released LFM3 on October 27, 2026. The first release covers six models of about 2B parameters, branched from LFM3-1.7B-Base: LFM3-1.7B (text), LFM3-VL-2B (vision, and it takes video input), LFM3-Audio-2B, LFM3-Omni-2B, and the decision models d1-2B and d1-Omni-2B. The text model has a 256K context, fits in about 1 GB at 4-bit, and is post-trained for agent harnesses such as Hermes Agent, Pi, and OpenClaw. It was pre-trained on 22T tokens. Liquid says every model can be fine-tuned on a single GPU. Weights are on Hugging Face. The models are also on Liquid's Playground and the Liquid AI API. The blog says day-0 support for llama.cpp, MLX, and vLLM. OpenRouter is listed as a launch partner. LFM3-Audio-2B supports 18 languages for speech: transcribe, synthesize, and translate. The weights are open under Liquid's LFM 1.0 license. The Hugging Face card shows license: other, license_name lfm1.0. Read the license for commercial terms. Last checked October 28, 2026.