Liquid AI released LFM3 on October 27, 2026, a family of six open-weight models around 2 billion parameters, all branched from one 1.7B backbone. The set includes a text model for local agents, a vision model that takes images and video, an audio model for speech in and out, an omni model that combines those inputs, and two decision models that return typed, calibrated answers instead of generated text. You pick the modality the job needs, fine-tune it on a single GPU, and run it on a phone, laptop, vehicle, or private server that already holds the data. The backbone is a tri-hybrid of gated short convolutions, sliding-window attention, and global attention, pre-trained on 22 trillion tokens. LFM3-1.7B has a 256K context window and is post-trained for agent harnesses, tool calling, and work inside browsers and business apps. LFM3-Audio-2B handles recognition, synthesis, and spoken tool calls across 18 languages. LFM3-Omni-2B is Liquid's first model that can see, listen, and speak from one checkpoint. The launch video shows that mix in use, from reading signs in another language and booking a flight from a spoken request to changing a molecular view by voice and adding a calendar invite. The decision models, d1-2B and d1-Omni-2B, answer routing, scoring, extraction, and pointing questions in one forward pass, including from images or audio, without writing a reply. Weights are on Hugging Face, in the Liquid Playground, and through the Liquid AI API. Day-0 support covers Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via AI Hub and GenieX, and Vercel, with Mercedes-Benz and Shopify involved in pre-release testing. Under the LFM Open License, commercial use is free below $10 million in annual revenue. Above that, a separate license is required. Liquid AI is a Cambridge company spun out of MIT CSAIL, founded by Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus.