Liquid AI launches CPU-friendly LFM2.5 encoders with 8K context

Released July 29, 2026, the 230M and 350M bidirectional encoders combine short convolution blocks and grouped-query attention and support 8,192-token inputs. Liquid AI reports about 3.7× ModernBERT-base’s single-pass CPU throughput. The 350M model averaged 81.02 across 17 tasks, placing fourth among 14 models. The weights are available on Hugging Face under LFM Open License 1.0.

Visit original source