Introducing LFM3: Made for specialized intelligence
Introducing LFM3
Made for specialized intelligence
Today, we’re introducing LFM3, a new family of compact, open-weight multimodal models made for specialized intelligence.
LFM2 was optimized for speed at the edge. LFM3 builds on that foundation with a redesigned architecture that delivers greater expressivity at a similarly compact scale.
What’s new in LFM3
Fine-tune 2B-class models for a specialized task on a single GPU
One shared backbone for text, vision, audio, omni, and decision models
A 256K context window; text model ~1 GB at 4-bit precision
New encoders for faster vision processing and streaming audio, plus video support
LFM3 gives you a more capable starting point for specialized applications ranging from real-time video understanding at the edge to on-device agents, while retaining the compactness and speed that made LFM2 practical.
Read the full LFM3 post for more details.
Start testing LFM3 in Liquid Console with included API credits.
© 2026 Liquid AI