LiquidAI releases LFM2.5-Encoders for fast long-context inference on CPU via Hugging Face

LiquidAI released two new encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, on Hugging Face as open-weight models. The models are designed for fast long-context inference on CPU, targeting document-scale tasks such as intent routing, policy linting, PII detection, and text classification.

The encoders are initialized from LiquidAI’s LFM2 decoder backbones and converted to bidirectional encoders using bidirectional attention masks, non-causal short convolutions, and masked language modeling with 30% token masking. Training proceeds in two stages: general language competence at 1,024-token context on a large web corpus, followed by long-context adaptation up to 8,192 tokens to strengthen factual, legal, and multilingual competence.

On benchmark tasks drawn from GLUE, SuperGLUE, and multilingual classification, LFM2.5-Encoder-350M ranks fourth among 14 models, with the three models ahead of it being substantially larger, including a 3.5B-parameter model nearly ten times its size. LFM2.5-Encoder-230M outperforms ModernBERT-base and all EuroBERT models while being smaller than most of them.

Inference speed measurements across full sequence lengths show the largest advantage on CPU, where LFM2.5-Encoder-230M is the fastest at every sequence length and about 3.7× faster than ModernBERT-base at 8,192 tokens (about 28 seconds versus over 90 seconds per forward pass). On GPU, the encoders lead for inputs longer than roughly 2,000 tokens.

The release includes CPU-only Hugging Face Spaces demos for zero-shot prompt routing, zero-shot policy linting, spell checking, PII detection across 16 languages, and a masked-diffusion text generation prototype. Developers can load and run inference with the transformers library and fine-tune the models for downstream tasks such as classification, token classification, regression, or retrieval.