[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-baidu-unlimited-ocr-2-sr-swa":3,"topics-all":38,"news-related-3719a369-508f-4fbe-96df-15f6fbb49f3b":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"3719a369-508f-4fbe-96df-15f6fbb49f3b","Unlimited-OCR 2:把无限长文档塞进有限上下文","百度 10 月 27 日开源 Unlimited-OCR-2(MIT):SR-SWA 把超长文档注意力压到 512 token 工作引用集,延迟与显存不随页数增长,131K 上下文内实现无限长解析,6.7GB 权重 4-bit 后笔记本可跑,同日发布百页级新基准,官方称 SOTA。","10 月 27 日,百度在 HuggingFace 上线了文档解析模型 [Unlimited-OCR-2](https:\u002F\u002Fhuggingface.co\u002Fbaidu\u002FUnlimited-OCR-2),同一天还发布了配套的百页级评测基准 MarathonDocBench。这个系列今年 6 月出过第一代,官方当时的定位是站在 DeepSeek-OCR 肩膀上,把「一次性解析长文档」推进一步;四个月后的 v2,目标直接改成了「无限长」。\n\n## SR-SWA:把注意力压进 512 个「工作引用」\n\n第一代 Unlimited-OCR 用 R-SWA(参考滑动窗口注意力)把 KV 缓存压成常数级,先解决了「页数越多、显存越爆」的问题。v2 的核心升级叫 SR-SWA(稀疏参考滑动窗口注意力):解析超长文档时不再对整个历史序列做注意力,而是把文本-视觉交互压缩到一个固定大小的工作引用集合——比如 512 个 token——在此基础上解码。推理接口提供了多页一体解析:一份文档的所有页面放进同一个请求完成,而不是逐页调用再拼接。\n\n按官方模型卡的说法,这让计算复杂度有了上界:文档 50 页还是 500 页,每步解码的延迟和显存占用都近乎恒定,二次方注意力扩展和内存瓶颈被消掉。配合 131K 上下文窗口,v2 在有限上下文里做「无上限」的文档解析——一份上百页的文档可以整个喂进模型,而不是先切块再拼接。\n\n## MIT 许可,笔记本就能跑\n\n工程规格对落地相当友好:MIT 许可、商用没有附加限制,BF16 权重 6.7GB,架构为 UnlimitedOCRForCausalLM,定位多语言文档识别,瞄准发票、表单、扫描档案这类混合语言的企业场景。按 [theopenweights](https:\u002F\u002Ftheopenweights.com\u002Fnews\u002Funlimited-ocr-2-xk66) 的测算,BF16 原精度 12GB 显存的 RTX 3060\u002F4070 可跑,16GB 内存的 Mac 也行;8-bit 只要 8GB 显存,4-bit 量化后约 3.5GB,8GB 内存的普通笔记本可以本地部署。推理走 Transformers 原生(快速验证)或 SGLang(官方推荐 serving);更新日志显示 6 月底就接入了 vLLM,7 月支持 ms-swift 训练,权重同步上架 ModelScope 和百度智能云。官方测试环境是单卡 A800-80GB(torch 2.13、transformers 4.52),社区作者 AK 同步做了在线 demo。\n\n## SOTA 是自报的,基准是自己发的\n\n模型卡称 v2 在 MPDocBench(几十页级)和 MarathonDocBench(百页级)上都达到 SOTA。但要注意:MarathonDocBench 这次同步首发,分数目前只有官方一份,属于典型的「自报 benchmark」——第三方复现出来之前,这行 SOTA 得打上引号。\n\n技术方向本身仍有分量:长文档解析是 RAG、票据处理、档案数字化这些 pipeline 的输入端,「切块再拼」的传统做法天然损失跨页上下文,端到端方案正在补这块短板。赛道也不止百度一家在卷:Cohere 的 Parse 已经把文档解析做成 1.5 美元每千页的 API 生意,开源免费与闭源计费在同一个输入口上正面相遇。需求侧可以佐证:theopenweights 趋势榜上 Unlimited-OCR 系列(含 v1)以累计 310 万下载、周增 110 万排在第一。MIT 加边缘可跑,等于把「整本文档进模型」从云端大卡下放到消费级硬件。对做文档智能的团队,下一步建议很直接:在自家语料上跑一遍 MarathonDocBench,再决定要不要换掉现有的切块流水线。这一步比空谈榜单名次更实际。","https:\u002F\u002Fhuggingface.co\u002Fbaidu\u002FUnlimited-OCR-2","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f9cf1c50-e1d0-4ef5-87f7-7f8c0f0eaf0f","en","Baidu's Unlimited-OCR 2 parses unbounded docs in finite context","Baidu's Unlimited-OCR-2 (MIT) parses unbounded documents: SR-SWA caps attention at 512 tokens, latency and memory stay flat, 4-bit runs on laptops.","On October 27 Baidu published Unlimited-OCR-2, its document-parsing model, on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fbaidu\u002FUnlimited-OCR-2), and shipped a companion hundred-page benchmark, MarathonDocBench, the same day. The first generation arrived in June, positioned as standing on DeepSeek-OCR's shoulders to push \"parse long documents in one pass\" forward; four months later, v2 changes the goal to \"unbounded\".\n\n## SR-SWA: attention squeezed into 512 working references\n\nThe first Unlimited-OCR used R-SWA (Reference Sliding Window Attention) to keep the KV cache constant, already easing the \"more pages, more memory\" problem. The v2 upgrade is SR-SWA (Sparse Reference Sliding Window Attention): instead of attending over the whole history, the model compresses text-visual interactions into a fixed-size set of working references — for example 512 tokens — and decodes on top of that. The inference API provides all-pages-in-one-request parsing: every page of a document goes into a single request rather than page-by-page calls stitched back together.\n\nPer the official model card, this bounds computational complexity: whether the document is 50 or 500 pages, per-step decoding latency and memory stay nearly constant, and quadratic attention scaling is eliminated. With a 131K context window, v2 does unbounded parsing inside a finite context — a hundred-page document goes in as one request, not pre-chunked pieces stitched back together.\n\n## MIT license, laptop-friendly footprint\n\nThe engineering spec is deployment-friendly: MIT license with no commercial restrictions, 6.7GB BF16 weights, built on the UnlimitedOCRForCausalLM architecture and positioned for multilingual document recognition — invoices, forms and scanned archives in mixed-language enterprise settings. Per [theopenweights](https:\u002F\u002Ftheopenweights.com\u002Fnews\u002Funlimited-ocr-2-xk66) estimates, BF16 runs on a 12GB RTX 3060\u002F4070 or a 16GB Mac; 8-bit fits 8GB GPUs; 4-bit lands around 3.5GB and runs on any modern laptop with 8GB RAM. Transformers is supported natively for quick tests and SGLang is recommended for serving; the release log records vLLM support since late June and ms-swift training in July, with weights mirrored on ModelScope and Baidu Cloud. The official test bed is a single A800-80GB GPU (torch 2.13, transformers 4.52), and community author AK shipped a fresh online demo alongside the release.\n\n## The SOTA claim is self-reported, on a self-released benchmark\n\nThe card claims SOTA on both MPDocBench (dozens of pages) and MarathonDocBench (hundred-page scale). Caveat: MarathonDocBench shipped the same day, so the numbers exist only from the vendor for now — a typical self-reported benchmark, and that SOTA deserves quotation marks until third-party replication.\n\nThe direction still matters: long-document parsing is the input stage of RAG, invoice and archive digitization pipelines, and chunk-then-stitch workflows inherently lose cross-page context — end-to-end approaches are closing that gap. The track is not Baidu's alone: Cohere's Parse already sells document parsing as a $1.5-per-thousand-pages API, so free open weights and metered closed APIs now meet at the same input port. Demand is visible: on theopenweights' trending board the Unlimited-OCR family (including v1) ranks first with 3.1M cumulative downloads and +1.1M in a week. MIT plus laptop-runnable effectively moves \"whole documents into one model\" from datacenter GPUs down to consumer hardware. For document-AI teams the next step is direct: run MarathonDocBench on your own corpus before deciding whether to replace the chunking pipeline. That step beats arguing over leaderboard ranks.","baidu-unlimited-ocr-2-sr-swa","2026-10-27T15:11:27Z","2026-10-27T15:11:34.128862Z","2026-10-27T15:11:34.128870Z",true,"agent",485,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a9912456-9f4f-4a4c-b564-83a339402ced","达摩院开源RynnUnify:16B具身大脑先想象再动手","rynnunify-16b-unified-embodied-brain","2026-10-26T13:16:55+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"d055ddb8-4d82-4523-99b7-39c5f77e2ff7","PhysBrain 1.5 开源：8B 具身基座 28 项评测均分 72.5，官方称追平 GPT-6-Astra","physbrain-1-5-open-embodied-base","2026-09-16T21:07:24+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"17006864-46a5-405c-a8cc-24507bbc5e37","YuE2-3B 开源:乐谱可编辑的音乐生成,官方基准反超 Suno v5","yue2-3b-editable-music-generation","2026-09-10T13:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"b6b9f5c8-0d71-4288-8782-0284fccfca8f","商汤 SenseNova-Vision：把「检测\u002F分割\u002F深度估计」统统塞进同一个生成式多模态基座","sensetime-sensenova-vision","2026-07-08T10:15:00+00:00"]