Baidu open-sources Unlimited OCR 2 for hundred-page documents
A 3.3-billion parameter vision-language model holds parsing accuracy on documents up to 300 pages, under the MIT licence, by capping attention to a fixed working set instead of scaling with document length.
Baidu open-sourced Unlimited OCR 2 this week, a vision-language model built to parse documents hundreds of pages long in a single request, released under the MIT licence with weights on Hugging Face.
The model uses what Baidu calls Sparse Reference Sliding Window Attention, which caps attention to a fixed 512-token working set instead of letting it grow with the document. On 100-page documents, Baidu’s own benchmarks put per-step decoding at 7.0 times faster than full attention and 5.6 times faster than its prior sliding-window approach, with accuracy holding up to 300 pages. On MarathonDocBench, a benchmark aimed at hundred-page documents, it scores 77.1 against 33.7 for Kimi-K3. The model has 3.3 billion parameters and runs in BF16 on a single GPU, with weights totalling around 6.7GB.
# weights and code
huggingface.co/baidu/Unlimited-OCR-2Why it matters: document OCR pipelines usually hit a wall well before a hundred pages, as attention cost scales with length and either latency or accuracy gives way. Open, MIT-licensed weights that hold up at that scale mean self-hosting the OCR stage of a document pipeline, rather than chunking files or paying a hosted API per page, becomes realistic.
The caveat: the speed and benchmark figures are Baidu’s own, not independently verified, and testing was done on a single Nvidia A800 GPU.