Baidu's Unlimited-OCR adds vLLM support and tops Hugging Face trending

Baidu’s Unlimited-OCR, a 3-billion-parameter open document-parsing model, added vLLM inference support on June 28, 2026 and ranked as the top-trending model on Hugging Face that day. The MIT-licensed model collected 1,800 GitHub stars within 24 hours of its June 22 launch.

The draw is what the architecture does with long documents. Most vision-language models see memory balloon as page count grows; Unlimited-OCR uses a technique it calls Reference Sliding Window Attention to keep the key-value cache constant regardless of document length, according to Baidu’s model card. That lets it parse an entire book in a single inference pass within a 32,768-token context, rather than chunking the document.

The model is a Mixture-of-Experts design with 3 billion total and about 500 million active parameters, built on top of the DeepSeek-OCR foundation model. It ships in two configurations, “gundam” for 640-pixel images and “base” for 1024-pixel images, and now runs on Transformers, vLLM, SGLang and Docker, per the project’s GitHub repository. The vLLM addition, contributed by the community, is what makes the model practical for higher-throughput serving.

Baidu’s own numbers should be read with care. The company reports about 93 on the OmniDocBench benchmark and roughly 6 percentage points higher accuracy than the leading open-source baseline at about 12 percent faster throughput. Those figures are company-supplied and not independently verified; on the third-party ParseBench, the model posted a mean score of 46.17 and a text-content score of 86.81.

The broader pattern is small Chinese open models competing on document AI through efficiency rather than raw scale. Whether Unlimited-OCR’s trending spot converts into sustained adoption will depend on independent benchmarks holding up.