DeepSeek is exploring a new architecture again, open-sourcing DeepSeek-OCR 2.

DeepSeek has released and open-sourced its next-generation document recognition model, DeepSeek-OCR 2. Utilizing the DeepEncoder V2 architecture, it upgrades traditional fixed-sequence image scanning to a semantic reasoning mode with causal attention. Through a lightweight language model that dynamically rearranges visual tokens, AI can understand complex documents (such as tables and multi-column layouts) in a logical order, much like humans. It achieved a record-breaking 91.09% overall score in the OmniDocBench benchmark, reducing reading order recognition error by 33%.