Unlimited-OCR
A newer find replaces this one: baidu/Unlimited-OCR-2. Kept here for reference.
What it is
Unlimited-OCR is a vision-language model from Baidu designed for parsing ultra-long documents without losing context. It utilizes Sparse Reference Sliding Window Attention (SR-SWA) to compress attention interactions to a fixed size, enabling constant-latency inference regardless of document length.
The problem it solves
Standard OCR and LLMs struggle with documents exceeding their context window, requiring complex chunking strategies that often break cross-page context and increase latency.
What you could use it for
- Extracting structured data from hundreds of pages of technical manuals
- Parsing long legal contracts or financial reports in a single pass
- Digitizing extensive academic archives or historical documents
Why it matters
Eliminates the quadratic scaling bottleneck of current LLMs on long documents, making it feasible to process entire books or massive datasets with consistent speed and memory usage.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.