Model Card: Baidu Unlimited OCR 2

🌏 中文版

Model Info

Highlights

• SR-SWA (Sparse Reference Sliding Window Attention): compresses cross-page text-visual attention into a fixed 512-token working reference set (a 128-token local window plus top-16 selection over 32-token blocks), keeping per-step decoding latency and memory footprint nearly flat regardless of how many pages a document has
• Scores 77.1 on MarathonDocBench (hundred-page documents), more than double the next-best model Kimi-K3 at 33.7; that's only 7.5 points below its own 84.6 on MPDocBench (dozens of pages)
• Against its own predecessor, Unlimited-OCR 1.0: on the same MarathonDocBench test, the predecessor scored 0.0 while this generation scores 77.1 — the earlier model collapsed outright on hundred-page documents; this one is the first to actually hold up as documents get longer
• 3.34B parameters (SAM ViT-B + CLIP-L encoder, 12-layer MoE decoder), runs on a single NVIDIA A800-80GB with BF16 weights around 6.7GB, and costs just 273 image tokens per page in multi-page mode

Benchmark Results

⚠️ All figures are Baidu's own self-reported results, published only as a leaderboard chart image on the model card (transcribed here by third-party coverage, not a data table). MarathonDocBench is a benchmark Baidu built itself; the dataset isn't publicly released, and no other team has independently reproduced these numbers yet.

vs. Predecessor and Competitors

Against predecessor Unlimited-OCR 1.0, MPDocBench only improved by 22.7 points (61.9 → 84.6), but MarathonDocBench jumped from 0.0 straight to 77.1. That gap says a lot: the predecessor's infer_multi could already feed an entire document into one request, but it never actually solved the computational complexity of long-sequence attention, so it fell apart past some page count. SR-SWA's fixed-size reference set is what actually fixes the bottleneck — not a bigger context window or a multi-page inference interface.

Against Qwen3.8-Max, currently the runner-up on MPDocBench, the two are nearly tied on short documents (84.6 vs. 78.3, a 6.3-point gap). But move to the hundred-page MarathonDocBench and Qwen3.8-Max crashes from 78.3 to 22.9 — a 55.4-point drop — while Unlimited OCR 2 loses only 7.5 points. A high score on a short-document benchmark doesn't guarantee stability on long documents; what actually determines long-document parsing quality is whether the attention mechanism was designed for extreme sequence lengths, not how many parameters a model stacks or how large a context window it advertises.

On licensing, Unlimited OCR 2 sits in the same open-weight camp as DeepSeek-OCR-2 (Apache 2.0) and PaddleOCR-VL-1.6 (Apache 2.0): weights, code, and the technical report are all released under MIT, with no separate paid API product line.

What This Means for Agent Development

• If you're building agents that process long documents (clause-by-clause contract comparison, annual financial-report archiving, RAG indexing for entire books): Unlimited OCR 2's pitch is "no chunking first" — infer_multi ingests a whole document in one request and preserves cross-page table, heading, and paragraph structure, which especially helps downstream tasks that need the full document structure (auto-generated tables of contents, precisely locating a cited passage) rather than page-by-page isolated text
• If you're running a document-parsing pipeline and want to self-host inference: 3.34B parameters on a single A800-80GB with roughly 6.7GB of BF16 weights means a mid-tier GPU is enough, so you don't need to call an external OCR API — a realistic option for data that can't leave your premises (finance, healthcare documents)
• If your task needs high accuracy on heading or chapter structure (auto-generating a table of contents or chapter summaries): note that heading hierarchy is the model's weakest sub-score (only 48.2 on hundred-page documents) — that part still needs human review or extra post-processing, not full automation
• Not a fit: scenarios that need an official hosted API with SLA guarantees — Baidu currently only offers open weights and Baidu Cloud documentation, not a standard managed API like OpenAI's or Anthropic's, so you'll need to evaluate your own deployment and operations cost

Today's Insight

I used to assume "long-document OCR" mainly meant chunking and stitching, or just brute-forcing it with a bigger context window. Unlimited OCR 2's MarathonDocBench result breaks that assumption: the predecessor used the exact same "ingest every page at once" strategy (infer_multi), and its score collapsed to zero anyway. Only switching to a mechanism like SR-SWA — one that compresses attention into a fixed-size reference set — kept parsing quality from falling apart as page count grew. The reminder: the bottleneck in long-sequence tasks is often not "is the context window big enough," but whether the attention mechanism's own computational complexity has actually been solved.

References

• Baidu: Unlimited-OCR-2 model card (Hugging Face)
• Baidu: Unlimited-OCR source code on GitHub
• Baidu Inc. on X: Unlimited OCR 2 is here
• MarkTechPost: Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing
• explainX: Unlimited-OCR-2 — Baidu's 3B MIT OCR Model, How to Run It
• AlphaSignal: Baidu Open-Sources Unlimited OCR 2 to Parse Hundreds of Pages Without Breaking