Liquid AI releases LFM2.5-Encoders: small models that classify long documents on a laptop CPU in under 30 seconds

Classification Work Doesn't Need a Frontier LLM. Liquid AI's LFM2.5-Encoders Show Why.

Every minute, across thousands of production pipelines, a generative LLM is doing something it was overqualified for: classifying a support ticket, flagging a Social Security number buried in a contract, deciding which department handles an incoming email. These are the highest-volume jobs in applied AI. Most teams route them through a per-token API because it was the easiest path to ship. Liquid AI's new LFM2.5-Encoders, released July 28, are built on a different premise: the relentless, high-volume tasks don't need a generative model at all.

The two open-weight models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, are bidirectional encoders designed for classification, token-level tagging, and retrieval 1. Both support an 8,192-token context window (roughly 13 to 15 pages of text) and are built to run on commodity CPUs without a GPU 1.

The deployment case lives in the latency gap. At 8,192 tokens, ModernBERT-base takes over 90 seconds for a single forward pass on CPU. The 230M-parameter encoder from Liquid AI does the same work in about 28 seconds, roughly 3.7 times faster 1. For a team running thousands of document classifications per hour on hardware they already own, that gap determines whether the pipeline holds under load or stalls.

Accuracy holds for the size. Across 17 tasks from GLUE, SuperGLUE, and multilingual benchmarks, tested against 14 encoder models, LFM2.5-Encoder-350M ranks fourth 1. The three models ahead of it are all larger, including a 3.5-billion-parameter model nearly 10 times its size. The 230M model beats ModernBERT-base and every EuroBERT variant while being smaller than most of them 1. All scores are means across five held-out seeds after full supervised fine-tuning, and the evaluation framework and raw results are open-sourced 1.

But the benchmark ranking is not the story. The story is what those numbers mean for infrastructure budgets.

Consider the jobs Liquid AI targets: intent routing, policy linting, PII detection, content classification 2. These run constantly 1. Feeding each of those through a generative LLM API means paying per token for a model whose generative capability (the ability to produce novel text) goes entirely unused during classification. A 354.5-million-parameter encoder, fine-tuned once for the task, runs on a CPU you already own with no per-token cost and no network round trip 3.

The PII detection demo makes the niche concrete. A fine-tuned model spots and removes 40 categories of personal information across 16 languages, running entirely on CPU 2. That is a task where latency, privacy, and cost all push in the same direction: away from the cloud, toward a small model that runs where the data already lives.

Liquid AI frames these encoders as part of a broader architectural shift. In regulated industries like finance, healthcare, and legal work, documents are long and frequently cannot leave in-house infrastructure. A CPU-only encoder that classifies a full contract in a single forward pass is a compliance advantage, not just a cost saving 1. For edge and embedded systems, from automotive controllers to industrial devices, a compact encoder handles command classification on the hardware already present, with no spare GPU and no cloud dependency 1.

The models also work as a cost layer in front of a larger system. A small encoder triages and filters incoming requests so an expensive generative model runs only when it is genuinely needed 1. At scale, that tiering is where infrastructure savings compound.

There is a caveat. All performance and speed claims originate from Liquid AI's own benchmarking. The evaluation code and raw results are open-sourced, making independent reproduction possible, but the numbers have not been independently verified by a third party 1.

Both models are available under the LFM Open License v1.0 on Hugging Face, fine-tunable with the standard transformers library, and accompanied by CPU-only demos for routing, policy linting, spell checking, and PII detection 3.

The teams getting AI infrastructure costs under control are not the ones buying cheaper GPUs. They are the ones asking, for each task, whether a generative LLM was ever the right tool. For the highest-volume jobs in applied AI, the answer is increasingly no.