ModelScope launches Unlimited-OCR: 3.3B model for full-document parsing
Parses entire documents in one shot with 32K output tokens and MIT license.
Deep Dive
Unlimited-OCR performs full-document parsing instead of cropped-region OCR. It features a 32K output length for long OCR sequences, base and gundam image modes for different document layouts, and uses Transformers inference + SGLang serving with OpenAI-compatible streaming requests. It is built to push DeepSeek-OCR-style document parsing further.
Key Points
- 3.3B parameter multilingual OCR model for full-document parsing (not cropped regions)
- 32K output tokens support long OCR sequences in one shot
- Two layout modes (base/gundam) and OpenAI-compatible streaming via SGLang
Why It Matters
Open-source, MIT-licensed OCR that handles entire documents at once simplifies digitization and document AI workflows.