--- title: LFM2.5-Encoders for Fast Long-Context Inference on CPU description: Liquid AI has released two open-weight encoder models — LFM2.5-Encoder-230M and LFM2.5-Encoder-350M — designed for fast, long-context NLP inference on CPU.... canonical: https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: LFM2.5-Encoders for Fast Long-Context Inference on CPU | daily.dev og:description: Liquid AI has released two open-weight encoder models — LFM2.5-Encoder-230M and LFM2.5-Encoder-350M — designed for fast, long-context NLP inference on CPU.... og:url: https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8 og:image: https://api.daily.dev/og/posts/x49IlHNC8.png og:image:alt: LFM2.5-Encoders for Fast Long-Context Inference on CPU og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # LFM2.5-Encoders for Fast Long-Context Inference on CPU **[Hugging Face](https://daily.dev/sources/huggingface)** · 6 min read · 2 upvotes · 1 comments ## Summary Liquid AI has released two open-weight encoder models — LFM2.5-Encoder-230M and LFM2.5-Encoder-350M — designed for fast, long-context NLP inference on CPU. Built on the LFM2 architecture, they support 8,192-token contexts and are approximately 3.7× faster than ModernBERT-base at long context on CPU. The models are initialized from LFM2 decoder backbones and converted to bidirectional encoders via masked language modeling training. They match or outperform larger models on GLUE, SuperGLUE, and multilingual benchmarks. Use cases include intent routing, policy linting, PII detection, and text classification. Both models are available on Hugging Face and can be fine-tuned with the provided tutorial. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much faster is LFM2.5-Encoder-230M than ModernBERT-base at long context on CPU? LFM2.5-Encoder-230M is about 3.7 times faster than ModernBERT-base at 8,192-token context on CPU. ModernBERT-base takes over a minute and a half per forward pass at that length, while LFM2.5-Encoder-230M takes about 28 seconds, letting a full contract or long transcript be scanned in under 30 seconds on a laptop CPU. _Teams weighing CPU-only NLP stacks can track encoder speed comparisons like this one on daily.dev._ ### What context length do LFM2.5-Encoder-230M and LFM2.5-Encoder-350M support? Both LFM2.5-Encoder-230M and LFM2.5-Encoder-350M support an 8,192-token context, with latency that grows slowly as input length increases. They were trained in two stages: a short-context masked-language phase at 1,024 tokens, followed by long-context adaptation extending to 8,192 tokens across factual, legal, and multilingual data. _Developers choosing an encoder for long-document tasks can follow releases like this on daily.dev._ ### How does LFM2.5-Encoder-230M compare to ModernBERT-base and EuroBERT on benchmark accuracy? LFM2.5-Encoder-230M beats ModernBERT-base and every EuroBERT model on a combined benchmark of GLUE, SuperGLUE, and multilingual classification tasks, despite being smaller than most of them. LFM2.5-Encoder-350M ranks fourth among 14 evaluated models, trailing only three larger models, including one nearly 10 times its size. _Anyone comparing encoder model sizes against accuracy can keep up with results like these on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@kartiknvj** · 0 upvotes > Converting decoder backbones into bidirectional encoders through masked language modeling is a smart way to reuse pretraining, and the CPU speedup at 8k context is what makes this practical for the boring high-volume jobs. The listed use cases (intent routing, PII detection, policy linting) are exactly the places where paying for a generative model per call never made sense. I would want to see the long-context latency curve past 4k tokens, since that is usually where encoder throughput on CPU starts to bend. ## Similar posts on daily.dev - [Up to 3.2x Faster Inference with LFM2.5-DSpark](https://daily.dev/posts/up-to-3-2x-faster-inference-with-lfm2-5-dspark-nwilr6rly) · Hugging Face · 0 upvotes · 0 comments - [Accelerating vision-language models with LFM2.5-VL-DSpark](https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6) · Hugging Face · 0 upvotes · 0 comments - [T5Gemma 2: The next generation of encoder-decoder models](https://daily.dev/posts/t5gemma-2-the-next-generation-of-encoder-decoder-models-nyxpiksgp) · Hacker News · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#nlp](https://daily.dev/tags/nlp), [#transformers](https://daily.dev/tags/transformers) [View this post on daily.dev](https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"LFM2.5-Encoders for Fast Long-Context Inference on CPU","url":"https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8"},"datePublished":"2026-07-28T15:02:02.553Z","dateModified":"2026-09-13T19:17:15.366Z","description":"Liquid AI has released two open-weight encoder models — LFM2.5-Encoder-230M and LFM2.5-Encoder-350M — designed for fast, long-context NLP inference on CPU....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d7a9bcb7ddb2053eed86b99a0b7cd419?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d7a9bcb7ddb2053eed86b99a0b7cd419?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"machine-learning,llm,nlp,transformers","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"LFM2.5-Encoders for Fast Long-Context Inference on CPU"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8","comment":[{"@type":"Comment","text":"Converting decoder backbones into bidirectional encoders through masked language modeling is a smart way to reuse pretraining, and the CPU speedup at 8k context is what makes this practical for the boring high-volume jobs. The listed use cases (intent routing, PII detection, policy linting) are exactly the places where paying for a generative model per call never made sense. I would want to see the long-context latency curve past 4k tokens, since that is usually where encoder throughput on CPU starts to bend.","datePublished":"2026-07-28T18:48:20.423Z","url":"https://daily.dev/posts/x49IlHNC8#c-QaePqIeaE","author":{"@type":"Person","name":"kartik-nvjk","url":"https://daily.dev/kartiknvj","image":"https://media.daily.dev/image/upload/s--3gGgsVCw--/f_auto/v1781456774/avatars/avatar_TvTVeiMdkRCqWUDullFmy?_a=BAMAMiWQ0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/lfm2-5-encoders-for-fast-long-context-inference-on-cpu-x49ilhnc8#faq","mainEntity":[{"@type":"Question","name":"How much faster is LFM2.5-Encoder-230M than ModernBERT-base at long context on CPU?","acceptedAnswer":{"@type":"Answer","text":"LFM2.5-Encoder-230M is about 3.7 times faster than ModernBERT-base at 8,192-token context on CPU. ModernBERT-base takes over a minute and a half per forward pass at that length, while LFM2.5-Encoder-230M takes about 28 seconds, letting a full contract or long transcript be scanned in under 30 seconds on a laptop CPU. Teams weighing CPU-only NLP stacks can track encoder speed comparisons like this one on daily.dev."}},{"@type":"Question","name":"What context length do LFM2.5-Encoder-230M and LFM2.5-Encoder-350M support?","acceptedAnswer":{"@type":"Answer","text":"Both LFM2.5-Encoder-230M and LFM2.5-Encoder-350M support an 8,192-token context, with latency that grows slowly as input length increases. They were trained in two stages: a short-context masked-language phase at 1,024 tokens, followed by long-context adaptation extending to 8,192 tokens across factual, legal, and multilingual data. Developers choosing an encoder for long-document tasks can follow releases like this on daily.dev."}},{"@type":"Question","name":"How does LFM2.5-Encoder-230M compare to ModernBERT-base and EuroBERT on benchmark accuracy?","acceptedAnswer":{"@type":"Answer","text":"LFM2.5-Encoder-230M beats ModernBERT-base and every EuroBERT model on a combined benchmark of GLUE, SuperGLUE, and multilingual classification tasks, despite being smaller than most of them. LFM2.5-Encoder-350M ranks fourth among 14 evaluated models, trailing only three larger models, including one nearly 10 times its size. Anyone comparing encoder model sizes against accuracy can keep up with results like these on daily.dev."}}]} ```