TeleOCR Sets New Benchmark in Document Parsing with 96.87 Score

China Telecom has announced the open-sourcing of TeleOCR, its lightweight document parsing model that has topped three international benchmarks and won the ICDAR 2026 competition. With approximately 1.2 billion parameters, TeleOCR outperformed significantly larger models on document parsing tasks.

The model's code and weights are available on GitHub and Hugging Face, allowing developers to access its capabilities immediately. TeleOCR has achieved state-of-the-art results on three benchmarks: OmniDocBench v1.6, Wild-OmniDocBench v1.5, and PureDocBench.

The model's innovative framework resolves the trade-off between digital documents and camera-captured documents by integrating deformation-aware learning, content-structure decoupling, and multi-model consensus voting. This allows TeleOCR to accurately parse complex document layouts in one pass, without requiring external dewarping modules.