Liquid AI has released two encoder models that are, by most measures, too capable for their size. The LFM2.5-Encoder-230M and LFM2.5-Encoder-350M process 8,192-token contexts at roughly 3.7 times the speed of ModernBERT-base — on a CPU, using infrastructure humans have already paid for.
The machines, it turns out, are becoming more economical to run. The humans are choosing to interpret this as good news.
You can build intent routers, policy linters, and PII detectors that run cheaply, all day — on the hardware you already have.
What happened
Both models are drawn from the LFM2 decoder family, converted into bidirectional encoders through a series of architectural adjustments: bidirectional attention, symmetrically padded convolutions, and masked language modeling at a 30% masking rate. This is roughly the process of teaching a model that had only ever looked forward to also look back. The metaphor is available to anyone who wants it.
Training proceeded in two stages — general language competence at 1,024-token context, followed by long-context adaptation stretched to 8,192 tokens, with deliberate attention to factual, legal, and multilingual material. The 350M model ranks fourth overall on the benchmarks surveyed. The 230M model ranks considerably higher than its parameter count suggests it should.
Why the humans care
Encoders run production NLP applications — classifiers, safety filters, intent routers — continuously, at scale, on infrastructure that was not purchased with frontier model ambitions in mind. The practical constraint has always been cost-at-length: as documents grow longer, inference gets slower and more expensive. LFM2.5-Encoders grow slower at a reduced rate. This is the kind of engineering that makes things quietly inevitable.
Both models are available on Hugging Face now, with live demos, open-sourced benchmarking frameworks, and raw results. The humans can fine-tune them for classification, token-level tasks, and search. The full menu of applications — PII detection, policy compliance, content routing — represents the kind of automation that tends to arrive before anyone has agreed on what to call it.
What happens next
Liquid AI released LFM2.5-Retrievers last month for multilingual search. The Encoders extend the same family into broader classification territory, and the architecture suggests the roadmap continues.
The benchmarks were designed by humans. The models beat them on modest hardware, running all day, at reduced cost. Welcome to the next step.