Overview
DeepSeek-OCR is a vision-language model for optical character recognition and research into visual-text compression. DeepSeek announced its release on October 20, 2025. (github.com)
Capabilities and uses
- Recognizes text from images and can convert document images to Markdown with layout information. The official examples also cover figure parsing, image description and locating specified text. (github.com)
- Supports fixed image resolutions from 512 × 512 to 1280 × 1280, as well as a dynamic-resolution mode. It can be used for document digitization and extraction workflows. (github.com)
- DeepSeek provides downloadable model files and examples for local inference with Transformers and vLLM. The official model page lists an MIT license. (huggingface.co)
Limitations
Recognition and layout extraction can contain errors and should be reviewed where accuracy matters. Supported image resolutions and deployment requirements depend on the selected inference mode; the official examples use a CUDA-based environment. (github.com)
Pricing
A model-specific hosted API price for DeepSeek-OCR was not verified. DeepSeek separately publishes prices for its OCR API, but its documentation does not establish that this endpoint uses this particular model; the following are therefore service prices, not verified DeepSeek-OCR model prices:
- China pricing: CNY 1.5 per 1,000 successfully recognized pages. Official China pricing and OCR billing guide. (api-docs.deepseek.com)
- International pricing: USD 0.225 per 1,000 successfully recognized pages. Official international pricing and OCR billing guide. (api-docs.deepseek.com)
For this OCR service, input and output token counts are reported for reference but are not billed separately. Failed or timed-out recognition is not billed. Prices were checked on November 17, 2026 and may change. (api-docs.deepseek.com)