--- title: Baidu Open-Sources Unlimited OCR 2 for Hundred-Page... description: Baidu has open-sourced Unlimited OCR 2, a 3.34B parameter vision-language model for parsing hundred-page documents in a single request. Its SR-SWA attention... canonical: https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing | daily.dev og:description: Baidu has open-sourced Unlimited OCR 2, a 3.34B parameter vision-language model for parsing hundred-page documents in a single request. Its SR-SWA attention... og:url: https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm og:image: https://api.daily.dev/og/posts/aOFBwYhhM.png og:image:alt: Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing 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. # Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing **[Machine Learning News](https://daily.dev/sources/mlnews)** · 5 min read · 1 upvotes · 0 comments ## Summary Baidu has open-sourced Unlimited OCR 2, a 3.34B parameter vision-language model for parsing hundred-page documents in a single request. Its SR-SWA attention scheme caps attention at a fixed 512-token reference set, keeping per-step decoding cost nearly flat as documents grow. It scores 77.1 on MarathonDocBench, well ahead of the next model Kimi-K3 at 33.7, while rival Qwen3.8-Max drops from 78.3 to 22.9 on long documents. Heading hierarchy is its weakest area at 48.2. Weights are MIT-licensed, runs on a single CUDA GPU in BF16, and all benchmark numbers are self-reported by Baidu. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How many parameters does Baidu's Unlimited OCR 2 model have and what license is it under? Unlimited OCR 2 has 3.34 billion parameters and ships under the MIT license, with weights on Hugging Face and code on GitHub. Active parameter count for the mixture-of-experts decoder is not disclosed. It runs on a single CUDA GPU in BF16, with Baidu testing on an NVIDIA A800-80GB; BF16 weights total about 6.7 GB. _daily.dev surfaces open-source OCR releases like this for teams evaluating document parsing stacks._ ### What is SR-SWA attention in Baidu's Unlimited OCR 2 model? SR-SWA, or Sparse Reference Sliding Window Attention, compresses text-visual attention into a fixed set of working references, with 512 tokens given as the example size. It pairs a 128-token local window with top-16 selection over 32-token blocks (16 blocks of 32 tokens). This keeps per-step decoding latency and memory nearly constant as document length grows, unlike full attention which scales quadratically. _developers scaling OCR to long documents can track attention-efficiency techniques like this on daily.dev._ ### How does Unlimited OCR 2 compare to other models on long document OCR benchmarks? Unlimited OCR 2 scores 77.1 overall on MarathonDocBench, a benchmark of hundred-page documents, versus 33.7 for the next-best model Kimi-K3 and 28.6 for Gemini-3.8-Flash. Qwen3.8-Max, which leads on the shorter MPDocBench at 78.3, collapses to 22.9 on MarathonDocBench, while Unlimited OCR 2 drops only 7.5 points from its 84.6 MPDocBench score. All figures are self-reported by Baidu. _teams choosing an OCR model for lengthy documents can weigh benchmark trade-offs like these via daily.dev._ ## Similar posts on daily.dev - [Unlimited-OCR: The Open-Source Model That Changed How I Think About Document AI](https://daily.dev/posts/unlimited-ocr-the-open-source-model-that-changed-how-i-think-about-document-ai-aqfn6tvjl) · Medium · 3 upvotes · 0 comments - [Qwen3-VL can scan two-hour videos and pinpoint nearly every detail](https://daily.dev/posts/qwen3-vl-can-scan-two-hour-videos-and-pinpoint-nearly-every-detail-5dvyilfvf) · Hacker News · 1 upvotes · 0 comments - [OCR Showdown — Evaluating Three OCR Models](https://daily.dev/posts/ocr-showdown-evaluating-three-ocr-models-ch9qm0xde) · Medium · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#open-source](https://daily.dev/tags/open-source), [#rag](https://daily.dev/tags/rag) [View this post on daily.dev](https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm) ```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":"Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing","url":"https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm"},"datePublished":"2026-10-27T06:52:02.946Z","dateModified":"2026-10-27T06:52:34.269Z","description":"Baidu has open-sourced Unlimited OCR 2, a 3.34B parameter vision-language model for parsing hundred-page documents in a single request. Its SR-SWA attention...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c2bb54332fbd7927d189a9f6a647f951?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c2bb54332fbd7927d189a9f6a647f951?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Machine Learning News","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":"Machine Learning News","logo":"https://media.daily.dev/image/upload/s--0PoEtdkd--/f_auto/v1750409401/squads/marktechpost?_a=BAMClqZW0","url":"https://daily.dev/squads/mlnews"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,open-source,rag","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Machine Learning News","item":"https://daily.dev/squads/mlnews"},{"@type":"ListItem","position":3,"name":"Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-aofbwyhhm#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Baidu's Unlimited OCR 2 model have and what license is it under?","acceptedAnswer":{"@type":"Answer","text":"Unlimited OCR 2 has 3.34 billion parameters and ships under the MIT license, with weights on Hugging Face and code on GitHub. Active parameter count for the mixture-of-experts decoder is not disclosed. It runs on a single CUDA GPU in BF16, with Baidu testing on an NVIDIA A800-80GB; BF16 weights total about 6.7 GB. daily.dev surfaces open-source OCR releases like this for teams evaluating document parsing stacks."}},{"@type":"Question","name":"What is SR-SWA attention in Baidu's Unlimited OCR 2 model?","acceptedAnswer":{"@type":"Answer","text":"SR-SWA, or Sparse Reference Sliding Window Attention, compresses text-visual attention into a fixed set of working references, with 512 tokens given as the example size. It pairs a 128-token local window with top-16 selection over 32-token blocks (16 blocks of 32 tokens). This keeps per-step decoding latency and memory nearly constant as document length grows, unlike full attention which scales quadratically. developers scaling OCR to long documents can track attention-efficiency techniques like this on daily.dev."}},{"@type":"Question","name":"How does Unlimited OCR 2 compare to other models on long document OCR benchmarks?","acceptedAnswer":{"@type":"Answer","text":"Unlimited OCR 2 scores 77.1 overall on MarathonDocBench, a benchmark of hundred-page documents, versus 33.7 for the next-best model Kimi-K3 and 28.6 for Gemini-3.8-Flash. Qwen3.8-Max, which leads on the shorter MPDocBench at 78.3, collapses to 22.9 on MarathonDocBench, while Unlimited OCR 2 drops only 7.5 points from its 84.6 MPDocBench score. All figures are self-reported by Baidu. teams choosing an OCR model for lengthy documents can weigh benchmark trade-offs like these via daily.dev."}}]} ```