Liquid opens two 8K-context CPU encoders

Liquid AI released 230M- and 350M-parameter LFM2.5 encoders with 8,192-token context, and says the smaller model completes a full-context CPU forward pass in about 28 seconds versus more than 90 seconds for ModernBERT-base—roughly 3.7x faster—in its release benchmarks. Both models are downloadable for fine-tuning, including the 230M checkpoint on Hugging Face, but the vendor measurements still need reproduction on an operator’s own hardware and documents. Builders running classification, policy filters, or intent routing should file this beside Microsoft’s case for routing security work to small models: a CPU encoder can screen long inputs before an expensive generative model sees them.