Liquid AI has released two new encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, designed to provide high-quality bidirectional encoding with efficient scaling up to an 8,192-token context length.

  • Built on the LFM2 hybrid architecture, the models use a bidirectional attention mask and non-causal short convolutions to process full context.
  • Training involves a masked language modeling objective with a 30% mask rate, followed by long-context adaptation to strengthen factual and multilingual competence.
  • Benchmarks show LFM2.5-Encoder-350M ranks fourth among 14 models on GLUE/SuperGLUE tasks, while the 230M variant beats ModernBERT-base.
  • On CPU hardware, the 230M model is approximately 3.7x faster than ModernBERT-base at 8,192 tokens, making it suitable for edge devices and regulated environments.

These open-weight models allow developers to fine-tune for classification, intent routing, and safety filtering directly on CPUs or in low-latency GPU deployments.