Baidu Releases Open-Source Unlimited OCR 2 With Lower Long-Document KV Cache Usage

What happened

Baidu released and open-sourced Unlimited OCR 2 for continuous long-document parsing. Its technical report pairs sparse reference sliding-window attention with chunked decoding, reducing peak KV cache usage from 47GB to 1.77GB in a hundred-page-scale single-sequence test, excluding model weights and other memory costs. In a separate 16-way concurrency test, throughput rose from 360 to 1315 tokens per second; QbitAI also reports parsing an 884-page English edition of The Lord of the Rings in approximately 42 minutes and 56 seconds.

Read next

• More changes in today’s briefingMistral releases improved Mistral Large 4 open weights for coding and cybersecurity

Reported by 2 sources

The story so far

Baidu released and open-sourced Wenxin Unlimited OCR 2, making its code and model weights available on GitHub and Hugging Face and offering it through Baidu AI Cloud and Baidu’s AI Open Platform. Baidu says the model can parse documents without a page limit. QbitAI reports that the model’s technical report combines sparse reference sliding-window attention with chunked decoding, reducing peak KV cache usage from 47GB to 1.77GB in a hundred-page-scale single-sequence test. Those figures exclude model weights and other memory costs. In a separate 16-way concurrency test, throughput increased from 360 to 1315 tokens per second. QbitAI also reports that the model parsed an 884-page English edition of The Lord of the Rings in approximately 42 minutes and 56 seconds. Leiphone reports scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench, along with seven times faster single-step decoding than Full Attention on hundred-page documents using the model’s SR-SWA attention mechanism.