DeepSeek-OCR 2 Open‑Source: Enabling Human‑Like Image Reading with Causal Reasoning

DeepSeek released DeepSeek-OCR 2, a next‑generation optical character recognition system that shifts visual encoding from a fixed scanning paradigm to semantic causal reasoning, allowing AI to read images in a human‑like order.

The model adopts the DeepEncoder V2 architecture, discarding the traditional CLIP encoder and using the lightweight language model Qwen2‑500M as the visual encoder core. A "causal flow query" mechanism enables dynamic reordering of visual tokens, forming a two‑stage "encoder reordering + decoder parsing" 1‑D causal inference pipeline.

Architecturally, the encoder combines a visual tokenizer built on an 80‑million‑parameter SAM‑base with two convolutional layers (output dimension 896) and an LLM‑style encoding module that blends bidirectional attention with causal attention via a custom mask. The decoder retains the 30‑billion‑parameter MoE from DeepSeek‑OCR, activating roughly 5 billion parameters at inference, and operates on only 256‑1120 visual tokens—far fewer than the 6000+ tokens typical of comparable models.

In benchmarks, DeepSeek‑OCR 2 achieved a 91.09% overall score on OmniDocBench v1.5 (covering nine Chinese/English document categories), a 3.73% gain over its predecessor, and reduced the edit‑distance metric for reading‑order recognition from 0.085 to 0.057. With the same token budget, its edit distance (0.100) outperformed Gemini‑3 Pro (0.115). Production tests showed duplicate‑rate reductions from 6.25% to 4.17% on online log images and from 3.69% to 2.88% on PDF pre‑training data.

The training pipeline consists of three stages: (1) language‑model pre‑training of DeepEncoder V2 on 160 A100 GPUs with batch size 640 for 40 k iterations; (2) frozen visual tokenizer with joint encoder‑decoder optimization using four‑stage pipeline parallelism; (3) frozen encoder while fine‑tuning the decoder for 20 k iterations at double speed. Training data comprise OCR 1.0, OCR 2.0, and generic vision data (80% OCR), balanced via sampling and label refinement.

DeepSeek‑OCR 2 targets two main scenarios: an online OCR service that equips DeepSeek‑LLM with image/document reading capability, and a batch PDF processing pipeline for pre‑training data. Its approach of applying language‑model architectures to visual encoding suggests a path toward unified multimodal encoders that could eventually handle images, audio, and video within a single model.

Limitations remain; on dense newspaper layouts the edit distance exceeds 0.13 due to only 250 k training samples. The team plans to increase local crop counts and expand training data to address this. Overall, DeepSeek‑OCR 2 demonstrates that causal reasoning can restructure visual understanding, delivering both performance and efficiency gains while opening new avenues for multimodal AI.

GitHub: https://github.com/deepseek-ai/DeepSeek-OCR-2<br/>Model hub: https://huggingface.co/deepseek-ai/DeepSeek-OCR-2