DeepSeek OCR

Overview

DeepSeek OCR accelerates document intelligence with a Context Optical Compression Engine. It compresses high-resolution documents into lean vision tokens, then decodes them with a 3B-parameter mixture-of-experts model to deliver near-lossless text, layout, and diagram understanding across 100+ languages.

The architecture has two stages: Stage 1 is the DeepEncoder (~380M) that compresses visual context, and Stage 2 is the MoE Decoder (3B) that reconstructs text with high fidelity. A multimodal bridge connects vision to language, and the data pathway ensures robust multilingual document processing.

Feature stack includes the vision encoder, a mode selector for different document types, structured output for downstream workflows, and compliance considerations. Benchmark comparisons show DeepSeek OCR leading in multilingual document understanding.

Deploy locally with GPUs or call via API, and integrate into your document workflows. Use cases span scanned books and reports, technical diagrams and formulas, multilingual dataset creation, and document conversion apps. With MIT open-source licensing and API token economics, DeepSeek OCR makes enterprise-grade OCR accessible to everyone.

Core capabilities

Context optical compression

Compress high-res documents into lean vision tokens

MoE decoder

3B-parameter model for near-lossless text reconstruction

100+ languages

Multilingual document understanding

Vision encoder

Advanced image-to-text encoding

Structured output

Ready for downstream workflows

Mode selector

Optimized for different document types

Deploy anywhere

Local GPU deployment or API access

MIT open source

Free and open licensing

Use cases

Document OCR

Multilingual text extraction

Technical diagram parsing

Document conversion

Dataset creation