Liquid AI releases LFM2.5 CPU-efficient encoders: 230M/350M models beat larger alternatives on long-context tasks

Liquid AI released two general-purpose encoder models on Hugging Face: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Both models match or beat larger encoders on GLUE, SuperGLUE, and multilingual benchmarks while maintaining fast inference speed. The 350M model ranks fourth among 14 models tested, outperformed only by larger models including a 3.5B parameter alternative.

The models support 8,192-token context windows with latency that grows slowly as inputs increase. On CPU, the LFM2.5-Encoder-230M is 3.7x faster than ModernBERT-base at 8,192 tokens (about 28 seconds per forward pass versus 90+ seconds). The encoders are built from the LFM2 decoder family, converted to bidirectional attention with non-causal convolutions, and trained on masked language modeling for two stages: general language competence at 1,024 tokens, then long-context adaptation to 8,192 tokens.

The practical impact is that developers can now run document-scale classification, intent routing, policy linting, PII detection, and text extraction tasks on commodity CPU hardware without cloud APIs. Liquid AI provides CPU-only demos for zero-shot prompt routing, policy enforcement, and spelling correction, all running in browser-based Hugging Face Spaces.

For practitioners: these models reduce the infrastructure cost of high-volume understanding tasks (classification, extraction, scoring) that run all day. An encoder fine-tuned on custom tasks is smaller, faster, and far cheaper than a generative LLM while remaining accurate enough for production classification pipelines. This suggests a bifurcation of the frontier: large reasoning models for complex tasks, lightweight encoders for scale-efficient classification.