Liquid AI's LFM2.5 encoders punch above their size on CPU

Encoders are quietly having a good year, and the interesting part of it is what 'small' now means. Liquid AI released two new bidirectional encoder models built on its LFM2 hybrid architecture, at 230M and 350M parameters, and the pitch is that they hold their own against much larger encoders while running comfortably on a CPU.

The headline numbers, per the company's own reporting, are worth reading carefully. On a suite of 17 classification and understanding tasks drawn from GLUE, SuperGLUE and multilingual benchmarks, the 350M model ranked fourth overall, behind three larger encoders including one at 3.5B parameters, roughly ten times its size. The 230M model, the company says, outperformed ModernBERT-base and every EuroBERT variant it tested against. At full 8,192-token context on CPU, the 230M is clocked at about 28 seconds per forward pass versus roughly 90 seconds for ModernBERT-base, a 3.7x gap.

Why does any of this matter if you are not shipping an NLP pipeline? Because encoders are the workhorse of the boring, high-volume work: intent routing, PII detection, policy checks, first-pass content filtering before you spend money on a big generative model. The default for a lot of teams over the last two years has been to throw a hosted LLM at that work. If a 230M open-weight model that fits on a CPU can genuinely handle document-length inputs at reasonable speed, the cost math for that pipeline shifts, and so does the on-premise story for regulated buyers in finance, healthcare and legal.

The honest caveat is that these are self-published benchmarks with five-seed averaging on standard suites, not independent evaluations, and 'beat ModernBERT-base' is a specific claim about specific tasks. Real-world accuracy on your data can slide either way. The write-up also does not spell out license terms beyond 'unrestricted download, fine-tuning, and deployment,' which is the sort of thing worth reading on the actual model card before you build a product on it.

The direction is the part worth watching. Small, long-context, CPU-viable encoders with an open evaluation harness turn a class of high-volume AI work back into something a small team can own end-to-end, rather than rent by the token.