Liquid AI Launches LFM3 Models That Beat the Last Generation with Fewer Parameters

AI lab Liquid AI has launched its new LFM3 model family, covering text, vision, audio and decision-making, with six models in the first batch. The series uses a brand-new triple hybrid layer architecture and was pre-trained on 22 trillion tokens. The first products include LFM-1.7B for text, LFM3-VL-2B for vision, d1-2B for decisions, LFM3-Audio for audio, plus LFM3-Omni-2B and d1-Omni-2B for all-round tasks, all built on LFM3-1.7B-Base. The text model LFM-1.7B supports a 256K context window and can handle long documents and multi-turn chats. According to the company, it beats the previous LFM2.5-2.6B while using 35% fewer parameters. The smaller size means the model can run efficiently on phones and other edge devices.