Liquid AI released two open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, on July 28, 2026. The models process long-context document tasks up to an 8,192-token window and run on standard CPU hardware.
In CPU benchmarks published on Hugging Face, LFM2.5-Encoder-230M processed long text sequences roughly 3.7 times faster than ModernBERT-base. At the maximum 8,192-token context length, ModernBERT-base required over a minute and a half per forward pass on CPU, compared to about 28 seconds for LFM2.5-Encoder-230M. Liquid AI did not specify the exact CPU architecture used to record the 28-second result. On GPU hardware, ModernBERT-base maintains a speed advantage for sequences below 1,000 tokens, while the LFM2.5 encoders run faster beyond 2,000 tokens.
Engineers built the encoders by modifying the existing LFM2.5-230M and LFM2.5-350M causal decoder backbones. The team replaced causal attention with bidirectional attention masks, added symmetrically padded non-causal short convolutions, and applied a 30 percent masked-language objective. Training took place in two stages, beginning with a short-context 1,024-token pre-training phase before extending to the full 8,192-token context window.
In benchmark evaluations across 17 tasks from GLUE, SuperGLUE, and multilingual benchmarks, LFM2.5-Encoder-350M placed fourth among 14 tested models, behind three larger systems including a 3.5-billion-parameter model. LFM2.5-Encoder-230M outperformed ModernBERT-base and all EuroBERT models tested. Liquid AI also released five CPU-only Hugging Face Space demonstrations, including a personally identifiable information detector that covers 40 data types across 16 languages.
