Baidu has open-sourced Unlimited OCR, a 3-billion-parameter vision-language model designed for efficient long-document parsing. Released on June 22, 2026, the model reads extensive documents in a single pass by maintaining a flat key-value (KV) cache, enabling superior handling of lengthy text without the memory overhead typical of transformer-based architectures.
Key Features
- Flat KV Cache: Unlike conventional models that grow memory usage with input length, Unlimited OCR keeps the KV cache flat, drastically reducing resource demands for long documents.
- Single-Pass Processing: The model processes entire documents in one forward pass, eliminating the need for chunking or iterative reading, which improves both speed and accuracy.
- 3 Billion Parameters: Compact enough for practical deployment yet powerful enough for complex OCR tasks, balancing performance and efficiency.
- Open-Source Access: Released under an open-source license, allowing developers and researchers to integrate, fine-tune, or extend the model.
Implications for 2026
As document digitization accelerates across industries—from legal and finance to healthcare and education—Unlimited OCR addresses a critical bottleneck: the inability of standard OCR models to handle very long documents without splitting them into fragments. By solving the KV cache memory issue, Baidu's model sets a new benchmark for scalable OCR in real-world applications.
Technical Insight
In transformer-based models, the KV cache stores attention keys and values to accelerate generation. For long documents, this cache grows linearly with input length, often causing out-of-memory errors or performance degradation. Unlimited OCR's architecture flattens this cache, likely through novel attention mechanisms or compression strategies, enabling consistent memory usage regardless of document length. The exact implementation details are expected to be published in Baidu's upcoming technical report.
Availability
Unlimited OCR is available on Baidu's official repositories and can be integrated via major machine learning frameworks. Community contributions and further optimizations are anticipated as the model gains traction.
This release underscores a broader trend in 2026: efficient, open-source AI models that democratize access to advanced document intelligence capabilities.
via MarkTechPost