DeepSeek OCR 2
DeepSeek OCR 2 is a document extraction model with published weights and inference code. Its examples distinguish layout-aware Markdown conversion from plain text recognition.
The encoder changes the reading order
DeepEncoder V2 can reorder visual tokens according to image content. The research contrasts this with a fixed scan from the top left to the bottom right. The reordered sequence then goes to the language model for interpretation.DeepSeek OCR 2 paper and publication record
Crops affect the visual-token budget
The documented default assigns 144 visual tokens to each crop and 256 to the whole image. The repository includes PDF processing and accelerated inference guidance.DeepSeek OCR 2 model cardDeepSeek OCR 2 inference code
Practical guidance. A dense page and a simple letter can need different image treatment. Keep the source page alongside extracted text so a reviewer can inspect small numbers, merged columns and missing footnotes.
Use the model-specific inference path
The model card’s example loads custom model code, applies the image prompt and calls model.infer.DeepSeek OCR 2 model card
Practical guidance. A generic text-generation example is not a complete document pipeline. Include image preparation, page ordering, output parsing and error reporting in the integration.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | January 2026 |
| Access | Downloadable |
| Model identifiers | deepseek-ai/DeepSeek-OCR-2 |
| License | Apache 2.0.DeepSeek OCR 2 model card |
| Default visual input | Up to six 768 × 768 crops plus one 1024 × 1024 image.DeepSeek OCR 2 model card |
| Output modes | Document-to-Markdown with grounding, or plain OCR without layout.DeepSeek OCR 2 inference code |
Related models
- DeepSeek V4.1 Flash
DeepSeek · DeepSeek V4.1
Multimodal mixture-of-experts model with a causal encoder-decoder design and separate computation budgets for reading input and generating text.
Launch date API · Downloadable - Janus Pro 7B
DeepSeek · Janus Pro
The larger Janus Pro model combines image understanding and image generation through separate visual encoders and a shared language model.
Launch date Research releaseDownloadable