Liquid AI open-weights two CPU-fast 8K text encoders
LFM2.5-Encoder-230M and 350M bring an 8,192-token window to BERT-class models and cut CPU latency roughly threefold at full length.
Small models, long inputs
Liquid AI released weights on 29 July for two bidirectional text encoders, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, aimed at the unglamorous middle of most AI systems: retrieval, reranking, classification and filtering, where a full generative model is overkill.
Both models are built on the company's LFM2 hybrid backbone, combining gated short convolution blocks with grouped-query attention, and both accept 8,192 tokens — a long window for this size class, where 512-token limits were the norm until recently. They share a hidden size of 1,024 and a 65,536-token vocabulary covering 15 languages.
Rather than train from scratch, Liquid AI converted its existing LFM2.5 decoders into encoders with three changes: switching attention masking to bidirectional, making the short convolutions non-causal, and continuing training with masked language modelling at a 30% mask rate, double BERT's original 15%.
The numbers
Across a 17-task evaluation suite, the 350M model posts a mean score of 81.02, placing fourth among the 14 models compared; the 230M version scores 79.29, ahead of ModernBERT-base at 78.19 despite being smaller. The sharper claim is throughput on commodity hardware: at 8K tokens the 230M encoder completes a forward pass in roughly 28 seconds on CPU, against more than 90 seconds for ModernBERT-base — about a 3.7x gap, which matters when embedding pipelines run on the same machine as the application rather than on a GPU.
Both checkpoints are on Hugging Face under the LFM Open License v1.0, Liquid AI's own terms rather than a standard permissive licence, so teams should read the conditions before shipping commercially.
Why it matters
Encoder models are the quiet workhorses of production AI — every RAG stack, moderation filter and semantic search index is full of them — yet they have improved far more slowly than generative models, with ModernBERT serving as the default for years. Pushing a sub-400M encoder to 8K context while keeping CPU inference tolerable moves capability toward the edge, where privacy, offline operation and per-query cost, not benchmark leadership, decide what gets deployed.