LFM3 brings six multimodal models with about 2B parameters
🎧 Listen to this story
The new Liquid Foundation Models (LFMs) generation, called LFM3, brings six models with about 2 billion parameters derived from a shared backbone: text, vision, audio, multimodal and decision-making models. The lineup includes video input for the vision model and an omni model that combines text, images, video and audio. The LFM3 team says LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters.
The architecture was designed for agentic applications and edge devices. According to the publication, the text model takes about 1 GB at 4-bit precision, has a 256K-token context window and can be fine-tuned on users’ data on a single GPU. The base model was pretrained on 22T tokens, while the vision and audio models use new in-house encoders.
Why it matters · editorial interpretation
The lineup combines compact multimodal models with an architecture aimed, according to the LFM3 team, at agents and edge devices. If the specifications and performance claim hold, the smaller size and single-GPU fine-tuning could make local adaptation easier.
Sources
Research