Baidu released Unlimited OCR 2 on Hugging Face on October 27, 2026 (model card release notes), an open-source vision-language document parser aimed at multi-page and hundred-page OCR. Weights are listed under the MIT license. Hugging Face safetensors metadata reports 3,343,978,497 total parameters (about 3.34B); the model card UI labels the size as 3B.
Key Facts
| Details | |
|---|---|
| Release | October 27, 2026 on Hugging Face (baidu/Unlimited-OCR-2); also listed on ModelScope and Baidu Cloud per the card |
| License | MIT (Hugging Face card / API tag license:mit) |
| Parameters | About 3.34B total (HF safetensors metadata); card UI shows "3B params". Active MoE parameter count for v2 is not stated on the card |
| Core change | Sparse Reference Sliding Window Attention (SR-SWA): compresses text-visual attention to a fixed working reference set (card example: 512 tokens) so per-step decode cost stays nearly constant as documents grow |
| Serving | Transformers for quick tests; customized SGLang wheel in-repo for serving. Tested by Baidu on an NVIDIA A800-80GB |
| Benchmarks (self-reported) | Claims SOTA on MPDocBench (dozens of pages) and MarathonDocBench (hundred-page); MarathonDocBench was released with the model. Independent replication not yet cited |
| Prior version | Unlimited-OCR, June 2026; v2 paper PDF is in the HF repo; card citation still points at the v1 arXiv entry (2606.23050) |
Background
Unlimited OCR 2 continues Baidu's line of long-document OCR after Unlimited-OCR. The card says SR-SWA moves the product from long-horizon to "unlimited-horizon" parsing inside a finite context window by bounding attention to a fixed reference set instead of letting text-visual attention grow with document length. Multi-page inference (infer_multi) takes ordered page images in one request so the model can use cross-page context; single-page infer defaults to dynamic tiling ("gundam" mode).
MarkTechPost describes a DeepSeek-OCR-style layout (vision encoder into a MoE decoder with 64 routed experts, 6 active per token, plus 2 shared experts) and quotes Baidu chart scores such as 84.6 on MPDocBench and 77.1 on MarathonDocBench. Those chart numbers appear only as images on the model page; we treat them as Baidu's own claims, not independently verified.
What We Don't Know Yet
- Active (routed) parameter count for Unlimited OCR 2; some social posts said "3B / A0.5B MoE," which the HF card does not confirm
- Public dataset release details for MarathonDocBench beyond the card's announcement
- Third-party replications of the MPDocBench and MarathonDocBench scores
- Whether a separate v2 arXiv abstract will replace the v1 citation on the card