Liquid AI released two encoder models, LFM2.5-Encoder-230M and -350M, built for efficient long-context processing on CPU by converting decoder architectures into bidirectional encoders through a bidirectional attention mask and non-causal short convolutions. The models are trained in two stages: general language competency at a 1,024-token context followed by adaptation to an 8,192-token context. The 350M variant ranks fourth among 14 tested models despite its smaller size and runs roughly 3.7x faster than ModernBERT-base at maximum context length on CPU hardware.
