Liquid AI Releases Open-Weight Bidirectional Encoders LFM2.5-Encoder-230M and 350M

Liquid AI released two open-weight bidirectional encoders, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, based on the LFM2 hybrid backbone. Both support 8,192-token context. They are converted from LFM2.5 decoder backbones by replacing causal attention with bidirectional, making short convolutions non-causal, and training with a 30% masked language modeling objective. Training had two stages: general language competence at 1,024 tokens, then context extension to 8,192 tokens. The 350M model achieved a 17-task mean of 81.02, ranking fourth among 14 models, behind XLM-R XL (3.5B), ModernBERT-large (395M), and XLM-R large (560M). The 230M model achieved 79.29, beating ModernBERT-base and all EuroBERT models. At 8K tokens on CPU, the 230M model runs one forward pass in ~28 seconds versus over 90 seconds for ModernBERT-base. The encoders are intended for edge devices, regulated on-premise systems, and high-volume pipelines. They load through transformers with trust_remote_code=True and require fine-tuning. Liquid AI also provided five CPU-only demos on Hugging Face Spaces.