On January 27, 2026, the DeepSeek team unveiled a technical report and made a new-generation OCR-specific model, DeepSeek-OCR 2, available as open-source software. Serving as an enhanced iteration of its forerunner, DeepSeek-OCR, this model marks a transformative leap in visual encoding, transitioning from static scanning to semantic reasoning through the introduction of the DeepEncoder V2 architecture. The crux of its innovation resides in the proposed 'Visual Causal Flow' mechanism. This mechanism empowers the model to dynamically rearrange processing sequences in accordance with image semantics, closely mirroring the logical flow of human reading. In the authoritative benchmark assessment OmniDocBench v1.5, DeepSeek-OCR 2 attained a holistic score of 91.09%, marking a 3.73% enhancement over its predecessor. Moreover, it achieved a 33% reduction in the edit distance metric concerning document reading order, underscoring a substantially improved capacity for comprehending logical structures. The model harnesses global visual information via a bidirectional attention mechanism and dynamically deduces optimal processing pathways using a causal attention mechanism. This enables it to efficiently condense intricate document content with a mere 256 to 1120 visual tokens. In real-world production settings, the model curtailed duplicate rates by 2.08% and 0.81%, respectively, when handling online user logs and PDF data, thereby showcasing its high practical utility. This upgrade not only elevates OCR performance but also substantiates the potential of language model architectures for visual encoding, thereby charting a technical course for unified multimodal encoders.