Liquid AI released two open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, built to process documents up to 8,192 tokens without a GPU. The company reports both models match or beat larger rival encoders on GLUE, SuperGLUE, and multilingual classification benchmarks, and that the 230M version runs about 3.7 times faster than ModernBERT-base at that context length on CPU. The release follows Liquid AI’s LFM2.5-Retrievers from the prior month, a search-focused model built on the same architecture.
An encoder is a different tool from a chatbot. It does not generate text; it converts text into a numerical representation that a classifier, router, or search index can score, and that representation work is where most production text volume actually sits. Spam filters, intent routers, PII detectors, and document classifiers run continuously in the background, not in response to a single prompt, and encoders are what power them.
Liquid AI built the new models by converting the causal decoder backbones LFM2.5-230M and LFM2.5-350M into bidirectional encoders. Each token now attends to neighbors on both sides instead of only the ones before it, and the company masked 30 percent of tokens during training. Training ran in two stages: a short pass at a 1,024-token context for general language competence, then an extension to the full 8,192-token window for factual, legal, and multilingual tasks.
On Liquid AI’s own evaluation of 14 models across 17 tasks drawn from GLUE, SuperGLUE, and multilingual classification, LFM2.5-Encoder-350M ranked fourth. The three models ahead of it were all larger, including one with roughly 3.5 billion parameters, nearly ten times the 350M model’s size. The smaller LFM2.5-Encoder-230M beat ModernBERT-base, along with every EuroBERT model tested, according to the company, while carrying fewer parameters than most of them. Liquid AI has open-sourced the evaluation framework and raw results, though the ranking itself remains the company’s own test.
The CPU gap widens as documents get longer. At 8,192 tokens, Liquid AI says ModernBERT-base takes roughly 90 seconds per forward pass against about 28 seconds for LFM2.5-Encoder-230M, the 3.7x figure cited above. On GPU the pattern flips at short inputs: ModernBERT-base leads below roughly 1,000 tokens, and Liquid AI’s encoders take over from about 2,000 tokens onward. These figures come from Liquid AI’s own benchmarking, not an independent lab, so the exact multiplier deserves a rerun on a team’s own hardware before it enters a capacity plan.
The operator implication sits in the CPU number, not the accuracy table. A retrieval-augmented generation pipeline typically splits into two halves: a generative model that writes the answer, and an embedding or classification layer that finds, ranks, and filters the source documents. If that second half runs acceptably on CPU at document length, a team can pull a GPU line item out of the retrieval side of its infrastructure bill, provisioning ordinary compute instead of accelerator capacity for the highest-volume, lowest-margin part of the stack. That only holds if Liquid AI’s parity claims survive contact with a team’s own documents and its own accuracy bar, since every comparison number here comes from the vendor that built the model.
Teams running document classification, PII detection, or long-context retrieval on GPU should benchmark LFM2.5-Encoder-230M or LFM2.5-Encoder-350M against their current stack before their next hardware renewal. A validated CPU swap changes both the cost and the deployment footprint of that layer, well before it touches anything on the generative side of the product.
Liquid AI published the LFM2.5-Encoder release on its Hugging Face blog on June 26, 2026.
