Unlimited OCR 2 Benchmarks: Where Baidu's Model Wins and Fails
Unlimited OCR 2 benchmark results on MP-DocBench and Marathon-DocBench, tested on handwriting, mixed languages, old prints, and 100-page documents.
What is Unlimited OCR 2 and why does it matter for long documents?
Unlimited OCR 2 is a document parsing model from Baidu built to convert scanned pages, PDFs, and photos into clean text, tables, formulas, and headings. The core problem it targets is simple to state and hard to solve: most OCR and vision-language models get slower and less accurate as documents get longer, because they try to hold the entire page history in memory while processing new pages. Baidu’s fix is an attention mechanism called SR SWA, which keeps a small, fixed set of reference points from earlier pages instead of re-reading everything. That keeps speed and memory roughly flat even as document length grows, which is the whole point of calling it “unlimited.”
TL;DR
- Unlimited OCR 2 holds around 77 out of 100 on the 100-page Marathon-DocBench test, while most competing models, including Qwen and Gemini variants, drop below 30 at that length.
- The previous Unlimited OCR version scored zero on the same 100-page benchmark, making this release’s long-document stability the headline improvement rather than a marginal upgrade.
- On shorter, dozens-of-pages documents (MP-DocBench), the model leads by only a few points over rivals, so its real edge shows up specifically at scale, not on short files.
- The SR SWA attention design trades full-context recall for a fixed set of “important” reference spots, which is why the model stays fast on long runs but can still drop details within a single page.
- Real-world tests found clear failure modes: it struggled with Urdu text mixed into an English document, misread decorative fonts and small print on an 1883 newspaper, and got a year and a price wrong on that same old print.
- Local testing showed roughly 8.3 GB of VRAM needed to run the model in full precision on a GPU, with a 2-second turnaround on a printed math problem.
- Clean printed English pages, including 50-plus page documents, came through with no dropped text in testing, suggesting the model is strongest on straightforward typeset material rather than handwriting or ornate fonts.
How does Unlimited OCR 2 perform on benchmark tests?
Baidu evaluated the model on two document benchmarks referenced in its GitHub materials: MP-DocBench and Marathon-DocBench. MP-DocBench tests documents that run to dozens of pages. On that benchmark, Unlimited OCR 2 comes out ahead of comparable models, including Qwen and Gemini, but the margin is a few points, not a wide gap.
Marathon-DocBench is the more revealing test. It pushes documents out to around 100 pages, which is where attention and memory costs typically pile up for vision-language models. On that benchmark, most tested models fall below 30 out of 100. Unlimited OCR 2 stays around 77. That gap is the entire value proposition of the SR SWA approach: rather than scaling badly as pages accumulate, the model’s accuracy curve stays mostly flat.
The comparison to its own predecessor is sharper still. The original Unlimited OCR reportedly scored zero on the 100-page benchmark, meaning it essentially failed outright on documents of that length. Version 2’s jump to 77 is less an incremental tweak and more a structural fix aimed squarely at the long-document failure mode.
What does SR SWA actually do, and what’s the tradeoff?
SR SWA stands for a sparse, reference-based sliding window attention approach. The plain-language version: instead of a model re-reading every prior page each time it processes a new one (which is how a lot of long-context vision models work, and why they get slower and hungrier for memory as documents grow), SR SWA keeps a small fixed set of reference points from what came before and largely ignores the rest.
Think of it as the difference between a student trying to hold an entire stack of notes in his head while writing (expensive, slow, and increasingly unreliable as the stack grows) versus a student who keeps only a handful of key reference points in mind and works from those. The second approach stays fast and uses consistent memory regardless of how many pages pile up. The tradeoff is that the model isn’t attending to literally everything at every moment, so details that fall outside its chosen reference points can get missed, even on a single page, not just across a long document.
That tradeoff shows up in testing: the model handled a 50-plus page document of plain English text without dropping a beat, but struggled with specific in-page details like decorative fonts, mixed scripts, or faint small print, areas where “look at less, but look at the right thing” doesn’t always pick the right thing.
Where does the model break: handwriting, mixed languages, and old prints?
Three failure categories showed up in hands-on testing with real documents, and they matter if you’re deciding whether this model fits your pipeline.
Mixed-language documents. A test image contained English, Urdu, and Indonesian text. The model read the English lines correctly, then got stuck in a repetitive loop once it hit the Urdu sentence it couldn’t parse. That points to a model trained primarily on English and Chinese rather than a broadly multilingual one. If your documents mix scripts, especially right-to-left or non-Latin alphabets, expect trouble.
Old or ornamental print. An 1883 newspaper page with decorative masthead graphics and antique fonts produced several concrete errors: the model skipped the large graphic masthead entirely, misread the year as 1893, and turned a price of “three pence” into something like “10 pounds.” Ordinary body text in the same document came through fine, read in under 2 seconds, but the ornamental and historical elements tripped it up. This suggests the model’s training data skews toward modern, clean typesetting rather than archival print styles.
Formula and table edge cases. On the positive side, a handwritten page of physics equations (special relativity, the photoelectric effect, blackbody radiation) was converted to LaTeX correctly in one pass. A printed math problem with a worked solution was parsed in about 2 seconds with every step of the simplification correct, though one fraction lost its visual fraction bar in rendering, a cosmetic issue rather than a numeric one. An invoice with table data converted cleanly overall, though one line item went missing, possibly due to context length limits during generation.
Is Unlimited OCR 2 worth using for document pipelines?
For a specific use case, long, clean, mostly-English or Chinese printed documents, it looks like a genuinely strong option. The 100-page benchmark result (77 versus sub-30 for rivals) is a meaningful jump, and real-world testing on 50-plus page plain-text documents backed that up with no dropped content. Running locally also means no per-page API cost and no document-length ceiling imposed by a vendor’s context window, assuming your hardware can handle the roughly 8.3 GB VRAM footprint for full-precision inference.
Where it’s a weaker fit: multilingual documents with non-Latin scripts, archival or heavily stylized print, and anything handwritten outside of clean structured content like equations. Financial documents and invoices look solid based on testing, but the missing invoice line is a reminder to verify table completeness rather than trust it blindly, especially on content near context-length boundaries. For any production use, spot-checking output against the source, particularly for dates, prices, and proper nouns, is still necessary.
Frequently Asked Questions
What is Marathon-DocBench testing?
Marathon-DocBench evaluates OCR and document AI models on very long documents, around 100 pages, to see whether accuracy holds up as page count increases rather than only testing short samples.
How much VRAM does Unlimited OCR 2 need to run locally?
Running the model locally in full precision used approximately 8.3 GB of VRAM during testing, loaded onto a GPU to process a single document image through the OCR pipeline.
Does Unlimited OCR 2 support multiple languages?
Testing suggests it handles English reliably and likely Chinese given its origin, but it struggled with Urdu mixed into a multilingual document, getting stuck in a repetition loop rather than parsing the text, indicating limited broader multilingual support.
How does Unlimited OCR 2 compare to its first version?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The improvement is largest on long documents: the original Unlimited OCR scored zero on the 100-page benchmark, while version 2 scores around 77, making long-document stability the primary upgrade over the prior release.
What kinds of documents cause the most errors?
Mixed-script text, decorative or antique fonts, and small ornamental print caused the clearest errors in testing, including a misread year and an incorrect price on an 1883 newspaper page, while clean modern printed text performed reliably.
