All Models
DeepSeek logo

DeepSeek-OCR

~3B

by DeepSeek

DeepSeek's ~3B MoE OCR model using optical context compression to encode full pages into compact token sequences. Outputs structured Markdown preserving text layout, tables, and mathematical formulas from images and PDFs.

Context WindowN/A
Parameters~3B
LicenseMIT
ModalitiesText, Image

Specifications

Technical details and pricing.

ProviderDeepSeek
Context WindowN/A
Release DateOct 1, 2025
ModalitiesText, Image → Text
CapabilitiesOCR, Document Parsing, Vision
LicenseMIT

Frequently Asked Questions

What is DeepSeek-OCR?

DeepSeek's ~3B MoE OCR model using optical context compression to encode full pages into compact token sequences. Outputs structured Markdown preserving text layout, tables, and mathematical formulas from images and PDFs.

What input formats does DeepSeek-OCR support?

DeepSeek-OCR accepts text, image as input and produces text output.

What is the context length of DeepSeek-OCR?

The context length for DeepSeek-OCR is not publicly documented.

Is DeepSeek-OCR open source?

DeepSeek-OCR is available under the MIT license.

Specifications are based on publicly available model documentation.