DeepSeek OCR

Description

DeepSeek OCR is a two-stage, transformer-based document AI that compresses page images into compact vision tokens before decoding them with a high-capacity mixture-of-experts language model. Stage one merges a windowed SAM vision transformer with a dense CLIP-Large encoder and a 16x convolutional compressor, while stage two uses the DeepSeek-3B-MoE decoder (~570M active parameters per token) to reconstruct text, HTML, and figure annotations with minimal loss.

Its core innovation is context optical compression: by reducing a 1024x1024 page to as few as 256 tokens, DeepSeek OCR enables long-document ingestion that would overwhelm conventional OCR pipelines, keeping global semantics while slashing compute. A mode selector ranges from Tiny (64 tokens) to Gundam (multi-viewport tiling), letting users tune the balance between speed and fidelity for invoices, blueprints, and large-format scans.

Trained on 30 million real PDF pages plus synthetic charts, formulas, and diagrams, it preserves layout structure, tables, chemistry SMILES strings, and geometry tasks, and it supports more than 100 languages across Latin, CJK, Cyrillic, and scientific scripts. Structured output can be emitted as HTML tables, Markdown, JSON, SMILES, and captions for direct ingestion into analytics pipelines.

The MIT-licensed weights (a ~6.7 GB safetensors checkpoint) can be run on-premises on local GPUs, avoiding cross-border data exposure, or called through DeepSeek's OpenAI-compatible API with token-based pricing. A single NVIDIA A100 can process roughly 200,000 pages per day.

DeepSeek OCR's Core Features

• Two-stage transformer OCR with context optical compression

• DeepEncoder combining windowed SAM and CLIP-Large with 16x compression

• DeepSeek-3B mixture-of-experts decoder (~570M active parameters)

• Mode selector from Tiny (64 tokens) to Gundam multi-viewport tiling

• Support for 100+ languages including CJK, Cyrillic, and scientific scripts

• Structured output as HTML tables, Markdown, JSON, SMILES, and captions

• MIT-licensed weights for on-premises local GPU deployment

• OpenAI-compatible hosted API with token-based pricing

How to use DeepSeek OCR?

1. 1

   Deploy locally: Clone the GitHub repo, download the ~6.7 GB safetensors checkpoint, and configure PyTorch 2.6+ with FlashAttention on a compatible GPU.

2. 2

   Or call the API: Use DeepSeek's OpenAI-compatible endpoints to submit images and receive structured text with token-based billing.

3. 3

   Choose a mode: Select Tiny, Base, Large, or Gundam to balance speed against fidelity for your document type.

4. 4

   Integrate outputs: Convert results to JSON, link SMILES strings to cheminformatics pipelines, or feed layout-aware HTML into automation and RAG workflows.

DeepSeek OCR's Use Cases

• Scanned books and reports
• Technical diagrams and formulas
• Multilingual dataset creation
• Document conversion apps
• On-premises OCR at scale

FAQ from DeepSeek OCR

• What is DeepSeek OCR?

  DeepSeek OCR is a two-stage transformer document AI that compresses page images into compact vision tokens, then decodes them with a mixture-of-experts language model to reconstruct text, HTML, and figure annotations for tables, charts, formulas, and diagrams with minimal loss.

• How does DeepSeek OCR compress long documents?

  It slices pages into patches, applies 16x convolutional downsampling, and forwards only 64 to 400 vision tokens to the MoE decoder, retaining layout cues while cutting context size significantly.

• How multilingual is DeepSeek OCR?

  It covers roughly 100 languages, spanning Latin, CJK, Cyrillic, and scientific notation, thanks to its extensive real and synthetic training data.

• Can DeepSeek OCR preserve tables and charts?

  Yes. Tests show near-lossless HTML and Markdown reproduction for tables and chart structures, enabling analytics pipelines without manual clean-up.

• Does DeepSeek OCR handle handwriting?

  Handwriting is not a core focus; performance is limited compared to specialized cursive OCR tools, so it is best paired with a handwriting engine when needed.

• Is DeepSeek OCR safe for regulated industries?

  Local deployment under the MIT license keeps data on-premises. When using DeepSeek's hosted API, users should consult compliance guidance regarding the company's cloud infrastructure.