LFM2.5-Encoders for Fast Long-Context Inference on CPU

Two new models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, have been released for fast, long-context inference, particularly on CPU hardware. These general-purpose encoders are adapted from LFM2 decoder backbones by enabling bidirectional attention and training with a masked language modeling objective. The models match or beat larger encoders on GLUE and SuperGLUE benchmarks while supporting an 8,192-token context. They demonstrate a significant speed advantage, with the 230M model being approximately 3.7 times faster than ModernBERT-base on long-context CPU tasks.