Liquid AI has released two new encoder models on Hugging Face: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. The company says the models are designed to match the quality of larger encoders while staying fast as inputs get longer.
The encoders support an 8,192-token context window and are intended for document-scale tasks that can run on everyday hardware, including CPUs. Liquid AI lists use cases such as intent routing, policy linting, personally identifiable information detection, and text classification. The post says the models are about 3.7 times faster than ModernBERT-base on long-context CPU inference.
This is a practical release for teams that need cheap, always-on text understanding rather than a large generative model. Encoders are often the quiet workhorses behind AI systems, helping route requests, filter content, or classify documents before a more expensive model is called. Faster small encoders can reduce both latency and serving cost.