deepseek/deepseek-ocr-2

DeepSeek-OCR 2 introduces a fundamental architectural rethinking compared to traditional OCR systems. Rather than relying on the standard CLIP visual encoder, the system swaps it out for a lightweight language model—specifically a Qwen2-0.5B component—paired with a causal flow query mechanism and the DeepSeek-3B-MoE decoder. This design shift moves document understanding from what DeepSeek calls "fixed scanning" to "semantic reasoning," processing image information based on meaning and reading order rather than raw positional patterns. The causal visual flow encoder allows the model to intelligently navigate document layouts by understanding spatial relationships through semantic context, enabling more human-like document interpretation.

The practical result of this architecture is dramatic efficiency: the model operates with 256 to 1120 visual tokens to cover complex document pages—approximately 80% fewer tokens than comparable approaches. On the OmniDocBench v1.5 benchmark, DeepSeek-OCR 2 achieves an overall score of 91.09%, marking a 3.73% improvement over its predecessor, while reading order recognition edit distance dropped from 0.085 to 0.057. The system outperforms Gemini 3 Pro on document parsing tasks. Released as an open-source model with frontier-level performance, it supports offline batch processing through frameworks like vLLM and SGLang, making high-quality document understanding accessible for developers building automated document pipelines, archival systems, or any application requiring accurate extraction from complex visual documents.