--- title: Build specialized multimodal models with LFM3 | daily.dev description: Liquid AI released LFM3, a new generation of Liquid Foundation Models built around one shared 2B-parameter-class backbone for text, vision, audio, omni, and... canonical: https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Build specialized multimodal models with LFM3 | daily.dev og:description: Liquid AI released LFM3, a new generation of Liquid Foundation Models built around one shared 2B-parameter-class backbone for text, vision, audio, omni, and... og:url: https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o og:image: https://api.daily.dev/og/posts/3KhcAOS3O.png og:image:alt: Build specialized multimodal models with LFM3 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. # Build specialized multimodal models with LFM3 **[Hugging Face](https://daily.dev/sources/huggingface)** · 12 min read · 0 upvotes · 0 comments ## Summary Liquid AI released LFM3, a new generation of Liquid Foundation Models built around one shared 2B-parameter-class backbone for text, vision, audio, omni, and decision tasks. A tri-hybrid architecture lets LFM3-1.7B outperform the previous LFM2.5-2.6B with 35% fewer parameters, while fitting in about 1 GB at 4-bit precision with a 256K context window. The base model was pretrained on 22T tokens. LFM3-1.7B scores 38.42 against 3.02 for its predecessor on WebArena-Verified and beats Qwen3.5-9B on long tasks in browsers and business apps. The vision model LFM3-VL-2B raises average vision benchmark scores from 57.6 to 71.4 versus LFM2.5-VL-3B. All models are open-weight on Hugging Face and fine-tunable on a single GPU. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does LFM3-1.7B compare to LFM2.5-2.6B in performance and size? LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters, thanks to a new tri-hybrid architecture combining gated short convolutions, sliding-window attention, and global attention layers. It matches LFM2-1.2B on short prompts per parameter and runs about 25% faster at 8K tokens of context. The text model fits in roughly 1 GB at 4-bit precision with a 256K context window. _Anyone picking a small on-device model for agentic tasks can track LFM3 benchmarks and releases on daily.dev._ ### What is the context window size of LFM3-1.7B and how was it trained? LFM3-1.7B-Base has a 256K token context window after pretraining on 22T tokens using the same 128K vocabulary as LFM2.5-8B-A1B. The final 3.5T tokens of training form a mid-training stage on curated data weighted toward agentic traces and tool calls, which is what extends the context window to 256K tokens. _Developers building long-context agent pipelines can follow model training details like this on daily.dev._ ### How does LFM3-VL-2B's vision encoder compare to SigLIP2 in efficiency? LFM3-VL-2B's in-house vision encoder uses 3 times less compute than the SigLIP2 encoder used in previous LFM VL models. The 164M-parameter encoder matches Qwen3.5-2B's vision encoder in accuracy while requiring 2.2 times fewer FLOPs, measured with the same LFM2.5-230M language backbone, and it halves temporal sequence length for video by embedding frame pairs into single slices. _Teams comparing vision encoders for edge deployment can keep tabs on efficiency benchmarks like this via daily.dev._ --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#local-ai](https://daily.dev/tags/local-ai), [#transformers](https://daily.dev/tags/transformers) [View this post on daily.dev](https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o) ```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":"Build specialized multimodal models with LFM3","url":"https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o"},"datePublished":"2026-10-27T17:19:22.514Z","dateModified":"2026-10-30T16:21:03.860Z","description":"Liquid AI released LFM3, a new generation of Liquid Foundation Models built around one shared 2B-parameter-class backbone for text, vision, audio, omni, and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ef197911f9e2c2f800e18b82d4174cfb?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ef197911f9e2c2f800e18b82d4174cfb?_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":0,"discussionUrl":"https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"deep-learning,local-ai,transformers","timeRequired":"PT12M"} {"@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":"Build specialized multimodal models with LFM3"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/build-specialized-multimodal-models-with-lfm3-3khcaos3o#faq","mainEntity":[{"@type":"Question","name":"How does LFM3-1.7B compare to LFM2.5-2.6B in performance and size?","acceptedAnswer":{"@type":"Answer","text":"LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters, thanks to a new tri-hybrid architecture combining gated short convolutions, sliding-window attention, and global attention layers. It matches LFM2-1.2B on short prompts per parameter and runs about 25% faster at 8K tokens of context. The text model fits in roughly 1 GB at 4-bit precision with a 256K context window. Anyone picking a small on-device model for agentic tasks can track LFM3 benchmarks and releases on daily.dev."}},{"@type":"Question","name":"What is the context window size of LFM3-1.7B and how was it trained?","acceptedAnswer":{"@type":"Answer","text":"LFM3-1.7B-Base has a 256K token context window after pretraining on 22T tokens using the same 128K vocabulary as LFM2.5-8B-A1B. The final 3.5T tokens of training form a mid-training stage on curated data weighted toward agentic traces and tool calls, which is what extends the context window to 256K tokens. Developers building long-context agent pipelines can follow model training details like this on daily.dev."}},{"@type":"Question","name":"How does LFM3-VL-2B's vision encoder compare to SigLIP2 in efficiency?","acceptedAnswer":{"@type":"Answer","text":"LFM3-VL-2B's in-house vision encoder uses 3 times less compute than the SigLIP2 encoder used in previous LFM VL models. The 164M-parameter encoder matches Qwen3.5-2B's vision encoder in accuracy while requiring 2.2 times fewer FLOPs, measured with the same LFM2.5-230M language backbone, and it halves temporal sequence length for video by embedding frame pairs into single slices. Teams comparing vision encoders for edge deployment can keep tabs on efficiency benchmarks like this via daily.dev."}}]} ```