LFM2.5-Encoders: Fast Long-Context CPU Inference

Liquid AI announced the release of LFM2.5-Encoders, two general-purpose bidirectional encoder models (230M and 350M parameters) initialized from LFM2.5 decoder backbones. They are pretrained with a masked-language objective (30% masking) over two stages: short-context (1,024 tokens) on a large web corpus and long-context adaptation (8,192 tokens) on a full data mix. On benchmarks, the 350M model ranks fourth among 14 models, beating larger ones up to 3.5B, while the 230M model outperforms ModernBERT-base and all EuroBERT models. Inference speed tests show LFM2.5-Encoder-230M is the fastest on CPU across all sequence lengths; at 8,192 tokens it processes a forward pass in ~28 seconds versus ~90 seconds for ModernBERT-base (3.7× faster). On GPU, the encoders lead for inputs above ~2,000 tokens. The models are open-weight on Hugging Face and support fine-tuning for classification, token-level tasks, and retrieval. Demos include zero-shot prompt routing, policy linting, spell checking, PII detection, and masked-diffusion text generation.