Baidu Releases Unlimited-OCR: A 3B Model for Long Document Parsing

Baidu has introduced Unlimited-OCR, a new 3B parameter model specifically engineered to enhance the processing of long documents. The model distinguishes itself by maintaining a flat Key-Value (KV) cache, a technical approach designed to optimize performance when parsing extensive textual and visual information.

Architectural Innovation in KV Caching

The core innovation of Unlimited-OCR lies in its management of the KV cache. By keeping the cache flat, the model addresses common bottlenecks associated with long-context document processing. This structural choice allows the 3B model to handle larger volumes of data more efficiently than traditional architectures that might struggle with memory overhead or latency during extended document analysis.

Efficiency and Scalability

With a parameter count of 3B, Unlimited-OCR balances computational efficiency with the capability to perform complex OCR tasks. The model is built to provide a scalable solution for users who require consistent performance across long-form documents. By focusing on the optimization of the KV cache, Baidu aims to streamline the parsing process, ensuring that the model remains responsive even when tasked with high-density document inputs.