Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU

Liquid AI just put out two open-weight encoders, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, and the pitch is refreshingly narrow: these aren't chatbots, they're the boring plumbing underneath classifiers, PII detectors, and safety filters that companies run constantly and rarely think about. Both handle an 8,192-token context, roughly 13 to 15 pages, so a single pass can chew through a full contract or a complete patient chart without chunking tricks.

What's clever is how they got made. Liquid AI didn't train encoders from zero. It took its existing LFM2.5 decoder backbones and performed surgery: swapped the causal attention mask for a bidirectional one, made the short convolutions non-causal with symmetric padding, and retrained with masked language modeling at a 30% mask rate, double BERT's original 15%. Training happened in two stages, starting at 1,024-token context on general web text, then stretching to the full 8,192 tokens with a mix leaning into legal, factual, and multilingual data.

The benchmark numbers back up the bet. Across 17 tasks pulled from GLUE, SuperGLUE, and multilingual sets, the 350M model landed fourth out of fourteen models with a mean score of 81.02, trailing only XLM-R XL at 3.5B, ModernBERT-large, and XLM-R large — all considerably bigger. The 230M version came in sixth, beating ModernBERT-base and every EuroBERT variant, including one that's nearly ten times its size. Liquid AI even ran its own methodology through Apache-2.0 open-sourcing, fixing the transformers version and using consistent AdamW settings so the comparison isn't skewed by dependency drift.

Speed is really the headline, though. On CPU at the full 8K context, the 230M model finishes a forward pass in about 28 seconds versus more than 90 for ModernBERT-base. That's the kind of gap that matters for a car's onboard system, an on-prem finance or healthcare pipeline that can't ship documents to the cloud, or any high-volume setup using a cheap encoder as a first filter before something bigger. Liquid AI shipped five CPU-only demos to prove it, including a PII detector spanning 40 categories across 16 languages and, oddly enough, a masked-diffusion chatbot that generates text by unmasking tokens iteratively.

Both models load through Hugging Face's transformers library with trust_remote_code enabled, and neither works out of the box for tasks — they're base encoders meant to be fine-tuned, with Liquid AI's own tutorial walking through long legal documents as the example use case. The choice between the two is uncomplicated: grab the 350M when accuracy is the priority, the 230M when you need more throughput or you're stuck on constrained hardware.