Liquid AI has released two open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, designed for efficient long-context inference on CPUs. These models support an 8,192-token context and are built by converting LFM2 decoder backbones into bidirectional encilers using masked language modeling.

  • The models match or beat larger encoders on GLUE, SuperGLUE, and multilingual tasks.
  • LFM2.5-Encoder-230M is approximately 3.7× faster than ModernBERT-base at 8,192 tokens on CPU.
  • Both models are available for fine-tuning on classification, routing, and extraction tasks.

The release aims to provide a cheaper and faster alternative to generative LLMs for high-volume production NLP applications running on standard hardware.