The mechanism is the release
Most document models read a file page by page, each page parsed in isolation, the results stitched together afterwards. Anything crossing a page boundary, a continuing table, a section started three pages earlier, is reconstructed by the stitching or lost. Unlimited OCR 2 skips the stitching: all pages go in as one request, cross-page context kept.
The reason that has been impractical is arithmetic. In ordinary attention, every token looks back at everything written so far, so memory and step time grow with the output. Baidu's answer, SR-SWA, caps the look-back: each step attends to a fixed working set of about 512 tokens, a local window plus a sparse selection of reference blocks. By limiting each decoding step to a bounded set of local and selected reference tokens, Baidu says SR-SWA keeps the attention-related decoding cost nearly constant as document length grows. That does not make total processing time, image encoding or overall memory requirements constant.
Read the card's slogan with that in mind: "true, unbounded document parsing within a finite context window". The window is 131,072 positions. At 273 image tokens per page, a hundred pages consume 27,300 of them before any output. Image encoding and total runtime still rise with page count. What Baidu has removed is the per-step penalty for length: a bottleneck, and narrower than unlimited suggests.
A public test only Baidu has run
On MPDocBench, a public benchmark of 433 documents and 3,246 pages, Baidu scores Unlimited OCR 2 at 84.6 overall, a narrower vendor-reported lead over Qwen3.8-Max at 78.3. On text the two are half a point apart, 94.8 to 94.3; on formulas, 89.4 to 89.2. A real lead, not a rout, and still the vendor's own scoring.
MarathonDocBench is the newer test: 100 documents of 100 to 300 pages each, 18,000 pages in total. Baidu published the dataset on October 26, the day before the model, with page images, annotations and evaluator files, scored through the MPDocBench-Parse toolkit on the same eight metrics. That makes the headline result open to independent testing, even though the published scores so far come from Baidu.
On that test, Baidu's figures put Unlimited OCR 2 at 77.1: Kimi-K3 scores 33.7, Gemini-3.8-Flash 28.6 and Qwen3.8-Max 22.9. The chart's most striking figure is the 0.0 for the previous Unlimited-OCR, but Baidu's technical report supplies the footnote the chart omits: the predecessor was not evaluated on MarathonDocBench because it supports inputs of at most 40 pages. The printed zero is therefore not a measured parsing score; it represents a model that was excluded from the evaluation.
The chart cannot say how much of the rivals' distance is parsing quality and how much is input handling (page limits, context handling, output format) on a 300-page single-request test. Only independent runs can settle that. The data and protocol are public; none outside Baidu has reported a run yet.
Until then, the sub-scores are the most useful part of Baidu's reporting, because they cut against the headline. Heading hierarchy, which tells a retrieval system where a paragraph belongs, scores 60.1 on MPDocBench and falls to 48.2 on the long documents this model exists for. Tables fall from 87.8 to 74.8. On Baidu's own scoring, heading hierarchy is substantially weaker than text and formula extraction on the long-document benchmark. The 48.2 heading-hierarchy score is a reason to test structural accuracy on real documents, not a directly interpretable failure rate for every deployment.
What to check before you build on it
The practical question is not whether the mechanism works. A fixed attention budget is a coherent way to stop per-step costs growing, the weights are MIT-licensed, and both benchmarks have publicly available evaluation materials, subject to their respective usage terms. The right test is your own long documents, scored on structure as well as text.
Three things deserve weight in that test. First, the tested hardware is an NVIDIA A800-80GB, a data-center card, even though the BF16 weights are only about 6.7 GB; what it needs at a hundred pages on a workstation card is not stated. Second, the recommended serving path is a customized Baidu build of SGLang, not upstream, serving documents in 10-page windows; Baidu's own note says a smaller window hurts cross-page merging, the feature the model is bought for. Third, multi-page mode encodes each page as a single global view, where single-page mode tiles for detail; small type and dense tables are what that view is likeliest to miss, and tables are already the second-weakest sub-score at length.
None of this cancels the release. Document parsing is unglamorous infrastructure many AI systems depend on, and a small, permissively licensed model holding a whole filing would matter if its accuracy held up under someone else's scoring. The claims to test are two: whether independent runs reproduce the gap behind the 77.1, and whether the structure survives the distance.
Sources
- [1] Baidu, "Unlimited OCR 2" model card, Hugging Face (primary source: release dated 2026/10/27, SR-SWA description, benchmark chart, inference and serving instructions, tested hardware)Read source
- [2] Baidu, Unlimited OCR 2 model config, Hugging FaceRead source
- [3] Baidu, "Unlimited OCR Works 2" technical report, Hugging Face (Tables 1–2 carry the full MPDocBench and MarathonDocBench scores; the Table 2 note records that the first model was not reported on MarathonDocBench)Read source
- [4] Baidu, MarathonDocBench dataset, Hugging Face (posted October 26, 2026: 100 documents of 100–300 pages, page images, ground-truth annotations and evaluator files)Read source
- [5] Baidu, Unlimited-OCR code repository, GitHub (MIT license file)Read source
- [6] MPDocBench-Parse paper, arXiv (the public multi-page benchmark: 433 documents, 3,246 pages)Read source