Article 332: Baidu's Unlimited-OCR-2 Is a 3B MIT-Licensed Model That Reads a Hundred-Page Document in One Pass - Because It Stopped Remembering Everything
Baidu released Unlimited OCR 2 today under MIT licence: a 3B vision-language model that parses an entire multi-hundred-page document in a single request, with per-step decoding latency and memory use that stay nearly flat as page count grows. Weights are on Hugging Face, mirrored to ModelScope, available as a managed Baidu Cloud API, with a public demo space.
The model is the smaller half of the story. The attention mechanism is the story.
SR-SWA: bounded memory on purpose
Standard attention is quadratic. Double the pages, quadruple the cost, and long before a hundred pages you hit a wall that no amount of GPU fixes. Baidu's answer is Sparse Reference Sliding Window Attention: instead of keeping the full text-visual attention history in memory, the model compresses it down to a fixed-size set of working references - on the order of 512 tokens - and slides that window forward as it reads.
It is not remembering the whole document. It is remembering a working summary of what it still needs, and letting the rest go. Constant cost per page, at the price of a bounded ability to look backwards.
That trade is exactly right for documents and exactly wrong for other things. A contract, a scan, a hundred-page filing, a textbook: the useful context is local, and cross-page tables and heading hierarchies matter more than paragraph forty-two. A codebase or a long conversation is the opposite case - the thing you need is often the thing that scrolled out of the window.
The benchmark is theirs
Baidu also published MarathonDocBench, a hundred-page ultra-long-document parsing benchmark, and claims SOTA on it and on MPDocBench for dozens-of-pages documents. Both are the company's own numbers, and one of the benchmarks is new and theirs.
To be fair, they also shipped the benchmark as a dataset rather than a leaderboard screenshot, which is more than most labs do on launch day. That is checkable by outsiders in a way a press release is not. It is still not independent verification.
Why this one matters more than it looks
The 3B parameter count and MIT licence put this in reach of a laptop-class GPU or a single modest cloud instance. Not a frontier lab. A records department, an insurance back office, a law firm with 400,000 pages of scanned filings and no appetite for sending them to an API. Document parsing is one of the few genuinely large, genuinely boring AI markets, and the ceiling on it was always cost per page. Open weights under MIT put a hard floor under that price.
Key Takeaway: This is the third wave of the same pattern - a Chinese lab ships a small, permissively licensed, narrowly excellent model, and the frontier labs' answer is that it is not a real benchmark. DeepSeek did it to reasoning. Qwen did it to general chat. Baidu is doing it to documents. The competitive pressure in open weights is not coming from where Western labs are looking, and constant-memory attention is now a solved problem in at least one narrow domain. Someone will generalise it.
Sources: Hugging Face, baidu/Unlimited-OCR-2 model card, 27 October 2026; GitHub, baidu/Unlimited-OCR (MIT); AlphaSignal, "Baidu Open-Sources Unlimited OCR 2 to Parse Hundreds of Pages Without Breaking," 26 October 2026; MarkTechPost, 26 October 2026