Unlimited-OCR
Document-parsing model published by Baidu on 22 June 2026 under an MIT license, 3 billion parameters. Its pitch is right in its stated motto, one-shot long-horizon parsing: instead of chaining one OCR pass per page, it exposes a multi-page inference call that swallows a whole PDF at once. A 32768 token window, BF16 safetensors weights, served through Transformers, vLLM (with a dedicated Docker image) or SGLang.
Strengths
- MIT license on a model of this size, still rare in the document-parsing category
- Native multi-page inference: the PDF is handled in one pass rather than stitched back afterwards
- Runs locally, so sensitive documents never travel to a third party
Limitations
- Self-hosted means hardware and a serving chain on you: the license is free, running it is not
- Recent model published in mid-2026, with little production feedback: validate it on your own corpus before leaning on it
- "A hundred pages in one pass" is a video promise, not a guarantee: the 32,768 token window holds both the page images and the returned text, so a dense document goes through in batches. Measure how many pages of your own contract fit in one request before building the pipeline on it ([Instagram reel DdJ-0WRQgmR](https://www.instagram.com/reel/DdJ-0WRQgmR/))
Best for
- Feeding a RAG with long PDFs without breaking reading order or tables
- Processing confidential documents you cannot send to an external API