Paper 2601.20552
DeepSeek-OCR 2: Visual Causal Flow
- Published
- Jan 2026
- Research lab
- DeepSeek
- Citations
- 53
- GitHub
- 3.2K stars
01 In brief
Summary
DeepSeek-OCR 2 introduces DeepEncoder V2, a novel vision encoder that replaces the CLIP component with a compact LLM (Qwen2-0.5B) to enable causal reordering of visual tokens, mimicking human visual scanning.
The encoder uses a dual attention mask: bidirectional for visual tokens and causal for learnable query tokens, allowing queries to attend to all visual tokens and preceding queries.
This design creates a two-stage cascade of 1D causal reasoning, aiming for genuine 2D understanding.
The model maintains a visual token budget of 256-1120 tokens per image via a multi-crop strategy.
On OmniDocBench v1.5, DeepSeek-OCR 2 achieves 91.09% overall, a 3.73% improvement over DeepSeek-OCR, with lower edit distances for text, formula, table, and reading order.
It also reduces repetition rates in production (from 6.25% to 4.17% for online logs, and 3.69% to 2.88% for PDFs).
Training involves three stages: encoder pretraining, query enhancement, and decoder specialization.
The architecture shows potential for unified omni-modal encoding and future improvements in 2D reasoning.
02 From the paper
Abstract
We present DeepSeek-OCR 2 to investigate the feasibility of a novel encoder-DeepEncoder V2-capable of dynamically reordering visual tokens upon image semantics. Conventional vision-language models (VLMs) invariably process visual tokens in a rigid raster-scan order (top-left to bottom-right) with fixed positional encoding when fed into LLMs. However, this contradicts human visual perception, which follows flexible yet semantically coherent scanning patterns driven by inherent logical structures. Particularly for images with complex layouts, human vision exhibits causally-informed sequential processing. Inspired by this cognitive mechanism, DeepEncoder V2 is designed to endow the encoder with causal reasoning capabilities, enabling it to intelligently reorder visual tokens prior to LLM-based content interpretation. This work explores a novel paradigm: whether 2D image understanding can be effectively achieved through two-cascaded 1D causal reasoning structures, thereby offering a new architectural approach with the potential to achieve genuine 2D reasoning. Codes and model weights are publicly accessible at http://github.com/deepseek-ai/DeepSeek-OCR-2.