Liquid AI releases LFM2.5-Encoders, open-weight encoder models built for fast CPU long-context inference

Liquid AI has released two new open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, aimed at classification and routing workloads that need to run continuously and cheaply on CPUs rather than GPUs. The company published the announcement on its official Hugging Face blog on July 28, 2026.

What's new

According to Liquid AI's own post, the release adds "two new encoder models on Hugging Face: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M," built for understanding tasks — classification, retrieval, routing — rather than text generation. Both are small by current standards: 230 million and 350 million parameters respectively. Despite the size, Liquid AI says the models "match or beat larger encoders on GLUE, SuperGLUE, and multilingual tasks," with the 350M variant ranking fourth among 14 tested models despite being smaller than most of the encoders it outperformed.

The headline capability is long-context efficiency on ordinary hardware. The models support "8,192-token context with latency that grows slowly as inputs get longer," and Liquid AI reports they run "about 3.7× faster than ModernBERT-base at long context" specifically on CPU infrastructure — a deliberate design target given how much production classification and routing work runs on CPU fleets rather than GPU clusters. Both models are "open-weight and available on Hugging Face today," free to download, fine-tune, and deploy.

Context

Encoder models occupy a different niche than the generative chat models that dominate headlines: they're the workhorses behind document classification, semantic search, spam and content filtering, and routing decisions inside larger AI systems — tasks that run at high volume and need to be cheap per-call rather than maximally capable. ModernBERT, the baseline Liquid AI benchmarks against, has been a widely used reference point for efficient encoders since its release, making a 3.7x CPU speedup a meaningful, concrete claim rather than a vague efficiency pitch. The release continues Liquid AI's focus on models sized and engineered for edge and resource-constrained deployment, rather than competing directly on frontier generative capability.

Why it matters

Most of the industry's attention goes to ever-larger generative models, but a large share of production AI spend sits in exactly the kind of high-volume, latency-sensitive classification and routing work these encoders target — and that work is usually run on CPUs for cost reasons, not GPUs. A small, open-weight encoder that is both more accurate than comparably sized alternatives and several times faster on CPU at long context is a direct cost lever for any team running document-scale classification in production. Because the weights are open, teams can self-host and fine-tune rather than pay per-call for a hosted API, which matters most for exactly the high-volume, always-on workloads this model is built for.