Liquid AI ships encoders that read both ways

Liquid AI released two new encoder models on July 28, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, bidirectional transformers derived from the company’s LFM2.5 decoder backbones. Unlike a causal decoder built to generate text, an encoder reads a document from both directions at once, which makes it better suited to classification, token-level tagging, and semantic search rather than generation.

Both models handle up to 8,192 tokens of context with manageable latency growth as input gets longer. On Liquid AI’s own numbers, LFM2.5-Encoder-350M ranks 4th among 14 tested models on a combined GLUE, SuperGLUE and multilingual benchmark, and the smaller 230M variant outperforms ModernBERT-base despite its size, processing a full 8,192-token input in about 28 seconds on CPU, roughly 3.7 times faster than ModernBERT-base at the same length. Liquid AI frames the practical use case directly: an entire contract or support-thread transcript scanned and classified in well under half a minute on ordinary laptop hardware, no GPU or outbound API call needed.

What it means for operators

Most teams doing document classification or retrieval default to a GPU-hosted encoder without checking whether a CPU-only alternative would clear their latency bar, especially for tasks like triaging long support threads or contracts where you don’t need generation, just accurate structure extraction. Liquid AI has built its whole LFM2.5 family around exactly this on-device, no-cloud-dependency pitch, in the same vein as other small models built for on-device use. A same-size encoder beating a widely used baseline like ModernBERT is the kind of vendor claim that’s cheap to verify yourself: pull both models, run them against your actual document set, and see if the CPU path is genuinely fast enough to retire a GPU dependency for that one workload.