LFM2.5-Encoder-350M

LFM2.5-Encoder-350M is an open-weight bidirectional encoder released by Liquid AI in mid-2026, converted from the LFM2.5-350M decoder backbone rather than trained from scratch. The team replaced causal attention with bidirectional masking and applied masked-language-modeling training, producing an 8,192-token-context encoder aimed at edge, regulated, and high-volume pipeline settings.

Liquid AI positions the encoder family for understanding and classification workloads: long documents, zero-shot routing, policy linting, and PII detection across 16 languages, all runnable on CPU-only hardware. Both encoder variants ship under the LFM Open License v1.0 with Hugging Face checkpoints; fine-tuning is required for task-specific use.