The family includes six models in the 2B-parameter range: a text model, a vision model that takes video input for the first time, an audio model, a first omni model and two decision models. The architecture is a "tri-hybrid" combining short convolutions, sliding-window attention and global attention. The text model was pre-trained on 22 trillion tokens, has a 256K-token context window and takes about 1GB at 4-bit precision.
According to the company, LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, with the largest gains on agentic tasks: on WebArena-Verified it scored 38.42 against 3.02. The LFM3-VL-2B vision model raised the average across 62 vision benchmarks from 57.6 to 71.4. The company also built new vision and audio encoders and says audio time to first token can be as low as 50 milliseconds.
All figures come from the company and have not been independently verified. The text model is post-trained to work in agent harnesses such as Hermes Agent, Pi and OpenClaw.