Liquid AI released two small open-weight text-understanding models, LFM2.5-Encoder-230M and 350M, on Hugging Face. They handle inputs of about 8,000 tokens and stay fast as text gets longer, running roughly 3.7 times quicker than ModernBERT-base on ordinary processors at that length. The models are general starting points and must be fine-tuned before they produce useful task output.
What changed
Comparable open text-understanding models such as ModernBERT slowed sharply as inputs grew, making long documents expensive to process on ordinary processors.
What it unlocks
Classifying, routing or scanning a full contract, transcript or support thread of up to roughly 8,000 tokens in under 30 seconds on a laptop processor, without a graphics card or an external API.
- 8,192-token context
- about 3.7x faster than ModernBERT-base at long context on CPU
- about 28s versus over 90s per forward pass at 8,192 tokens on CPU
- 230M and 350M parameter sizes
- 350M ranks 4th of 14 models across 17 tasks
What you need to act on it
- download the open weights from Hugging Face
- fine-tuning on the target task for anything beyond masked-token prediction
- the transformers library with trust_remote_code enabled
Sources