Liquid AI Releases Two Open-Weight LFM2.5 Encoders for Long CPU Workloads

The 230M and 350M LFM2.5 Encoders support 8,192-token inputs and CPU inference. Liquid AI reports roughly 3.7× lower long-context latency than ModernBERT-base.

What was released

Liquid AI released LFM2.5-Encoder-230M and 350M on Hugging Face on July 28, 2026. They convert LFM2.5 decoder backbones into bidirectional encoders using bidirectional attention, non-causal short convolutions, and masked-language training. Both support 8,192 tokens and downloadable weights for fine-tuning.

Intended workloads

  • Intent classification and prompt routing
  • Policy and PII detection
  • Document and support-thread classification or scoring
  • Token extraction and retrieval representations

Unlike a generative LLM, an encoder can cheaply classify or extract from an entire input. A small CPU model may be attractive for always-on filters or routers where cost and data locality matter.

Reported results

Liquid AI says it evaluated 14 models across 17 GLUE, SuperGLUE, and multilingual tasks with five held-out seeds. It reports that the 230M model beat ModernBERT-base and EuroBERT variants, and that an 8,192-token CPU pass took about 28 seconds versus more than 90 seconds for ModernBERT-base—about 3.7× faster.

Limits

These are developer-reported results under selected fine-tuning, hardware, and software conditions. The release article does not make every CPU and threading detail obvious. Teams should benchmark their own languages, documents, class imbalance, quantization or ONNX path, and PII false positives under equal budgets.

Official source