DeepSeekDeepSeek
DeepSeek-OCR Visual Context Compression

DeepSeek OCR

A picture is worth a thousand words. Discover how DeepSeek-OCR's visual modality compresses long text by 10x while preserving full semantic meaning.

DeepSeek-OCR Optical Compression

A 1000-word document needs ~1300 Text Tokens, but DeepSeek-OCR needs only ~100 Vision Tokens to reconstruct it perfectly. This 10x compression ratio means DeepSeek-OCR can process entire books at a fraction of the cost of traditional text tokenization.

DeepSeek-OCR DeepEncoder

A visual encoder designed specifically for compression within the DeepSeek-OCR pipeline. It combines SAM's local perception with CLIP's global semantics via a 16x compression layer, enabling DeepSeek-OCR to capture both fine-grained details and high-level meaning simultaneously.

DeepSeek-OCR All-Round Parsing

DeepSeek-OCR parses not just plain text, but complex charts, mathematical formulas, chemical molecular structures, and geometric shapes. This versatility makes DeepSeek-OCR suitable for scientific papers, financial reports, and technical documentation across diverse domains. The DeepSeek-OCR architecture handles all these modalities through a unified visual compression pipeline.

DeepSeek-OCR: Contexts Optical Compression Technology

Why use images for text? Because images are higher-dimensional information carriers. Drag the slider to see how DeepSeek-OCR maintains high accuracy with extremely low Token usage. The key insight behind DeepSeek-OCR is that visual representations encode spatial relationships, font hierarchies, and layout semantics that would require thousands of extra tokens to describe textually. DeepSeek-OCR fundamentally rethinks how language models consume document content.

DeepSeek-OCR Compression Ratio (Text/Vision Token Ratio)10x
No Compression (1x)DeepSeek Sweet Spot (10x)Extreme (20x)
Original

Pure Text

Cost: 1000 Tokens
Compress
DeepSeek-OCR

Visual Tokens

Cost: 100 Tokens
DeepSeek-OCR Accuracy98%

Optimal Zone

90.0%
DeepSeek-OCR Token Saving

DeepSeek-OCR achieves its remarkable compression by treating documents as images rather than character sequences. This optical approach preserves structural information like tables, headers, and formatting that traditional text tokenizers discard, while using 10x fewer tokens. The DeepSeek-OCR compression pipeline is trained end-to-end, ensuring that the visual tokens retain maximum semantic fidelity.

DeepSeek-OCR DeepEncoder Architecture

To achieve both High-Res Input and Low-Token Output, DeepSeek designed a serial architecture for DeepSeek-OCR called DeepEncoder. This three-stage DeepSeek-OCR pipeline processes documents at full resolution while aggressively compressing the output token count, balancing visual fidelity with computational efficiency.

Input Image
(High Res)

1. SAM Encoder

Visual Perception (80M)

2. Conv Compressor

16x Downsampling

3. CLIP Encoder

Visual Knowledge (300M)
Compressed
Latent Tokens

Hover or click on a module to see details

The DeepSeek-OCR architecture is deliberately modular: SAM handles perception, the Compressor handles efficiency, and CLIP handles understanding. This separation of concerns allows each component to be optimized independently, and the entire DeepSeek-OCR pipeline can be fine-tuned end-to-end for specific document types like invoices, academic papers, or handwritten notes.

多分辨率适配 (Adaptive Resolution)

Tiny
512x512
64 Tokens
Base
1024x1024
256 Tokens
Large
1280x1280
400 Tokens
Pro
Gundam
Tiling
Dynamic Tokens

DeepSeek-OCR 支持多种分辨率模式，从极速的 Tiny 模式到处理超大报纸的 Gundam (高达模式) 拼接模式，灵活应对不同场景。

FAQ

Traditional OCR only extracts raw text, losing important visual semantic information like layout structure, color coding, and chart relationships. Furthermore, feeding raw OCR-extracted text to LLMs consumes massive Tokens. DeepSeek-OCR compresses visual information directly into compact Tokens, saving cost while preserving the full semantic context of the original document.
Yes. DeepSeek-OCR uses an adaptive resolution strategy. For blurry or complex images, DeepSeek-OCR switches to high-res mode (or tiling), leveraging SAM's strong local perception capabilities to maintain high recognition rates even under challenging visual conditions.
The applications are enormous. DeepSeek-OCR excels at processing long financial PDF reports, analyzing complex scientific paper charts, parsing legal contracts, and enabling mobile AI assistants to 'see' your screen content without expensive cloud computation costs.
DeepSeek-OCR treats text as visual patterns rather than character sequences, which makes it inherently language-agnostic. The DeepEncoder architecture recognizes visual glyphs regardless of script — whether Latin, Chinese, Arabic, or Devanagari. This visual approach means DeepSeek-OCR handles multilingual documents with mixed scripts naturally, without needing separate language-specific OCR engines.
While GPT-4V and similar models process images at full token cost (often hundreds of tokens per image), DeepSeek-OCR is specifically optimized for text-heavy visual content compression. DeepSeek-OCR achieves 10x fewer tokens for document understanding by using its specialized DeepEncoder pipeline, making it far more cost-effective for document-processing workloads where the visual content is primarily textual.

Knowledge Quiz

How much have you learned?

QUESTION 1 / 4Score: 0

What is the core philosophy of DeepSeek-OCR?