Baidu open-sources Unlimited-OCR 2, delivering 5.6x faster single-step decoding for 100-page documents, eliminating the need to process entire pages at each step.

Beating AI Express (Dongcha) reports that Baidu has released and open-sourced Unlimited-OCR 2, a model with approximately 3 billion parameters, of which only around 500 million are activated during inference. The first version, launched in June, could parse around 40 consecutive document pages, but as page count grew, generating each token required attention to visual information from all pages, leading to progressively slower decoding. The second version introduces the SR-SWA sparse attention mechanism: instead of accessing visual data from all pages each time, the model focuses on the current page, selects a small amount of relevant visual information, and continues recognition in combination with recently generated text. In 100-page tests, single-step decoding is roughly 5.6 times faster than the previous R-SWA attention scheme and about 7 times faster than traditional full attention, with significant memory savings. In 100-page document trials, chunked decoding and cross-page cache reuse cut the peak KV cache memory for single-path inference from 47GB to 1.77GB. In another 16-path concurrent test, generation throughput jumped from 360 tokens per second to 1315 tokens per second. Baidu also built MarathonDocBench, a dataset of 100 documents ranging from 103 to 300 pages, to test long-document parsing capabilities. Unlimited-OCR 2 scored 77.06, outperforming Gemini 3.8 Flash (28.61) and Qwen 3.8 Max (22.93) in Baidu’s in-house benchmark tests. For extremely long documents, the system can use windowed processing and reuse part of previous window caches to further reduce memory overhead, though too small a window may harm cross-page content coherence. The paper also notes the new mechanism cannot fully retain a document’s global information, with overall structures like multi-level headings still prone to misidentification.