An encoder that reads thirteen pages at once, without a GPU

Liquid AI encoders read thirteen pages of text on a simple CPU. What these open-weights models change for data processing.

All the buzz goes to the models that write. Encoders, on the other hand, do not produce text: they read, and convert what they read into actionable representations to classify, route, or filter. They are the building blocks that run continuously behind a service, often on a simple CPU rather than a graphics card. Liquid AI is releasing two of them, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, with open weights.

The two models do not start from scratch. They build on the LFM2 hybrid architecture and convert a causal decoder, which only looks at what precedes, into a bidirectional encoder, where each token also sees what follows. Three modifications make this possible: the attention mask becomes bidirectional, short convolutions read from both sides thanks to symmetric padding, and the training masks 30% of the tokens where BERT masked 15%. A second phase then extends the context to 8,192 tokens, which is thirteen to fifteen pages swallowed in one go.

The advantage is especially visible on a CPU, and all the more so as the input grows longer. At full context length, the company measures about twenty-eight seconds per pass for the 230-million-parameter model, compared to over a minute and a half for ModernBERT-base, which is a factor of 3.7. On a graphics card, the gap narrows, with ModernBERT even keeping the advantage below a thousand tokens. On classification tasks, the 350-million model ranks fourth out of fourteen, behind three larger models, one of which is nearly ten times its size.

One thing sets this release apart from the crowd. The evaluation protocol is described—learning rate chosen on runs excluded from the final count, then averaged over five fresh runs—and the entire setup, including launchers and raw results, is open on GitHub. The models remain bases to be fine-tuned task by task. Demos running on CPU alone illustrate prompt routing, personally identifiable information detection, and text verification against rules written in natural language. On the sidelines, the encoder is repurposed into a conversational agent that composes its response by progressively unmasking, instead of writing it from left to right.