Liquid AI's 350M encoder beats 3.5B giants on CPU, costs grow slower with input

Liquid AI just dropped two open-weight encoder models that punch far above their weight class and are built for the real-world grind of CPUs, not just GPU benchmarks. The LFM2.5-Encoder-350M and its 230M sibling aren’t just smaller BERT alternatives; they’re a direct architectural argument that how you build the model matters more than how many parameters you stuff into it.

The 350M model scored an 81.02 average across 17 tasks from GLUE, SuperGLUE, and multilingual classification. That puts it fourth overall, but here’s the catch: the three models ahead of it are all significantly larger, with the top-ranked XLM-R XL being nearly 10 times its size at 3.5 billion parameters. Its smaller sibling, the 230M model, landed at 79.29, beating ModernBERT-base and every EuroBERT variant in the table, including a 2.1B model that scored a surprisingly low 72.19. “The gap is the stated reason Liquid AI built a general-purpose encoder instead of reusing the retrievers,” the team noted, pointing to the performance delta over their own retrieval-focused models.

The secret sauce is in the conversion from a decoder backbone. Liquid AI took their LFM2.5 decoders, ripped out the causal mask, made the convolutions bidirectional, and trained with a dense 30% mask rate—double BERT’s original 15%. The argument is straightforward: the hybrid architecture of gated convolutions and grouped-query attention scales more gracefully as inputs get longer. When you’re churning through a 15-page contract on a car’s onboard computer or a hospital server that can’t touch the cloud, that efficiency isn’t a nice-to-have.

Liquid AI shipped five CPU-only demos to prove it, covering everything from zero-shot prompt routing to a PII detector that handles 40 information types across 16 languages. The license is open, the training methodology is fully open-sourced under Apache 2.0, and the fine-tuning tutorial walks through long legal documents at the full 8,192-token context. For anyone building classifiers, safety filters, or intent routers that live on-device or behind a firewall, these models are a compelling reason to stop thinking of encoder architecture as a solved problem.