Liquid AI releases LFM2.5-Encoder models for efficient long-context CPU inference
Model release Provisional 95% confidence first seen
Liquid AI released two open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, designed for fast inference on CPUs with 8,192-token context windows. The models match larger models on benchmarks while achieving significantly faster CPU inference speeds, with the 230M variant processing 8,192 tokens in 28 seconds compared to 90 seconds for ModernBERT-base. These encoders enable cost-effective document classification and content filtering tasks on edge devices without GPU acceleration.