Liquid AI releases fast CPU-ready encoders for document-scale tasks

Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M as open-weight models on Hugging Face, extending the LFM2 family beyond retrieval into general-purpose text encoding. The encoders initialize from LFM2 decoder backbones and convert them to bidirectional models by applying a bidirectional attention mask so each token sees both left and right context, using non-causal short convolutions padded symmetrically to mix in neighboring tokens, and training with 30% token masking. Pretraining occurs in two stages: first, general language competence using masked-language modeling on a large web corpus at 1,024-token context; second, long-context adaptation extending to 8,192 tokens on the full data mix to strengthen factual, legal, and multilingual performance.

Benchmark results across GLUE, SuperGLUE, and multilingual classification tasks place LFM2.5-Encoder-350M fourth among 14 models tested, with only three larger models scoring higher—notably a 3.5B model nearly 10 times its size. LFM2.5-Encoder-230M beats ModernBERT-base and every EuroBERT model tested while remaining smaller than most. On CPU inference, the speed advantage is substantial: at 8,192 tokens, LFM2.5-Encoder-230M completes a forward pass in about 28 seconds, compared to over a minute and a half for ModernBERT-base—about 3.7× faster. LFM2.5-Encoder-230M is also fastest on CPU at every sequence length, even outpacing ModernBERT-base on short inputs. On GPU, LFM2.5-Encoder-350M and LFM2.5-Encoder-230M lead from about 2K tokens onward, whereas ModernBERT-base performs better below ~1K tokens on Apple GPU.

The article provides five live demos running in CPU-only Hugging Face spaces: zero-shot prompt routing (score prompts against custom routing lanes), zero-shot policy linting (score text against company rules), spell checking (token-level correction), PII detection (identifying and removing 40 kinds of personal information across 16 languages), and masked-diffusion text generation (an encoder-based chatbot that generates text by iteratively unmasking). For deployment, developers should choose LFM2.5-Encoder-350M when accuracy is paramount, or LFM2.5-Encoder-230M for tighter hardware constraints or higher throughput. Both models are available immediately on Hugging Face and can be loaded and run with the transformers library in a few lines of Python; they can also be attached to custom heads for fine-tuning on classification, token classification, regression, or retrieval tasks, with a tutorial provided for fine-tuning on long legal documents at 8,192-token context.

Liquid AI's new encoders address a persistent production-NLP need: running high-volume understanding tasks cheaply and fast on CPU. The company released LFM2.5-Retrievers for multilingual search last month; these encoders extend that family with a broader mandate, using masked-language pretraining instead of retrieval objectives so they can be fine-tuned for classification, token-level tasks, and search alike. The models inherit the fast-inference architecture of the LFM2 decoder backbone but flip it to bidirectional: each token sees neighbors on both sides, and short convolutions are padded symmetrically. Training happens in two stages—general language competence at 1,024 tokens on web text, then long-context adaptation extending to 8,192 tokens on a mixed corpus to strengthen factual, legal, and multilingual performance.

The benchmark results show LFM2.5-Encoder-350M ranks fourth among 14 models tested, with only three larger models ahead (including a 3.5B model nearly 10 times its size). LFM2.5-Encoder-230M outperforms ModernBERT-base on GLUE, SuperGLUE, and multilingual tasks despite being smaller. The speed advantage on CPU is dramatic: at 8,192 tokens, LFM2.5-Encoder-230M is about 3.7× faster than ModernBERT-base. On GPU, a similar pattern holds with a narrower margin—ModernBERT-base leads below ~1K tokens on Apple GPU, but LFM2.5-Encoders take the lead from about 2K tokens onward. For developers, this means scanning or classifying full documents (contracts, transcripts, long support threads) in under 30 seconds on a laptop CPU—a practical shift for always-on NLP applications.