Liquid AI released LFM3, a new open-weight model family. It covers text, vision, audio, omni, and decision models in the 2B parameter class. All models share a new tri-hybrid architecture. They run on any device and can be fine-tuned locally. The text model supports a 256K-token context window. LFM3-1.7B outperforms the prior LFM2.5-2.6B while using 35% fewer parameters. Models are available on Hugging Face, Liquid AI’s Playground, and its API. The family uses new in-house vision and audio encoders. Each model fits in about 1 GB at 4-bit precision. The weights are open and ready for deployment today.
What changed
LFM2 used a hybrid of short convolutions and attention with 3B-class models and shorter context windows.
What it unlocks
Running specialized 2B-class models for text, vision, audio, omni, and decision tasks on any device with fine-tuning on a single GPU.
- 2B parameter class models
- 22T tokens pre-training
- 256K context window
- 1.7B model outperforms 2.6B predecessor
What you need to act on it
- a single GPU for fine-tuning
- 4-bit precision for ~1GB memory footprint
Sources