Baidu open-sources Unlimited OCR for better long document parsing

Baidu has open-sourced the Unlimited OCR model, which has 3 billion total parameters but only activates 500 million during inference. It's designed to solve the slowdown issue in end-to-end OCR models when parsing long documents. The model uses a two-stage visual encoder and compresses a 1024x1024 PDF image into just 256 visual tokens.

Unlimited OCR was trained for 4,000 steps based on the DeepSeek OCR checkpoint, using about 2 million document samples on 8x16 A800 GPUs. In benchmarks, it scored 93.23 on OmniDocBench v1.5, beating DeepSeek OCR's 87.01 and DeepSeek OCR 2's 89.17. Its text edit distance is only 0.038, and it performs well on formulas and tables.

The model has already earned 6.8K stars on GitHub, marking a new step for efficient one-shot long document parsing.