Back to Blog

DeepSeek-OCR Released: High-Precision Document Understanding Vision-Language Model

DeepSeek-OCR is a vision-language model designed for visual text compression and document understanding, excelling at document-to-markdown conversion, OCR tasks, and chart parsing.

2025-10-23Product Release

Beyond Memory Anxiety: DeepSeek-OCR and the Vision of "Infinite" AI Context

Have you ever felt the frustration of an AI "losing its mind" halfway through a long document?. This is the "Long-Context Tax"—as text grows, the computational cost for LLMs explodes quadratically, leading to memory bottlenecks and high costs.

In a groundbreaking technical report, the DeepSeek-AI team presents a radical shift in perspective: DeepSeek-OCR. Instead of treating text as a massive string of data, they treat it as a visual medium that can be compressed, just like a JPEG image.

The Core Puzzle: Can a Picture Truly Save a Thousand Tokens?

The traditional "bridge" between seeing and reading in AI has been inefficient. Current models often fragment images into thousands of "vision tokens," slowing down the AI and hogging memory.

DeepSeek-OCR asks a provocative question: What is the minimum number of vision tokens needed to decode 1,000 words?. Their goal was to prove that visual modality is the ultimate compression medium for textual information.

The Secret Weapon: DeepEncoder’s "Variable Zoom"

To achieve extreme compression without losing data, the team built DeepEncoder (380M parameters). Imagine a high-end camera lens that can instantly switch between a wide-angle view and a microscope:

  • The Perception Eye (SAM-base): Uses "window attention" to capture fine details like the strokes of a character or the lines of a chart.

  • The Knowledge Brain (CLIP-large): Provides global context through dense attention.

  • The 16x Compressor: This is the magic. It bridges the two components, shrinking the mountain of visual data into a tiny "envelope" before the AI starts reading it.

DeepEncoder supports several Resolution Modes, allowing the AI to adapt its "vision" to the task:

  • Tiny/Small Mode: Uses as few as 64 to 100 tokens for simple slides or books.

  • Gundam Mode: A high-resolution tiling method for dense documents like newspapers.

The "Aha!" Moment: 10x Compression with 97% Accuracy

The results of this "optical compression" experiment are startling:

  • Near-Lossless at 10x: When the model compresses 10 text tokens into a single vision token, it maintains a decoding precision of 97%.

  • Functional at 20x: Even at a 20x compression ratio—where data is squeezed to the limit—the AI still recovers about 60% of the text correctly.

On the OmniDocBench benchmark, DeepSeek-OCR (using only 100 tokens) outperformed models like GOT-OCR2.0 which use 256 tokens. It also beat MinerU2.0 while using nearly 9 times fewer tokens.

Why This Matters: The "Biological Forgetting" Mechanism

This isn't just about reading faster; it's about how AI remembers. DeepSeek suggests that optical compression can simulate human memory decay.

  • Recent Context: Stored in high-resolution "Gundam" mode (crystal clear).

  • Distant History: Progressively resized into "Tiny" mode (blurry).

This allows for a "memory" that fades naturally over time, enabling theoretically unlimited context architectures where recent info is sharp and old info is compressed into a "vague recollection" that saves massive computing power.

Deep Parsing: Beyond Just Words

DeepSeek-OCR isn't just a reader; it’s a deep-parsing assistant. It can handle:

  • Complex Charts: Converting visual trends into structured HTML tables.

  • STEM Fields: Recognizing chemical formulas and converting them to SMILES format.

  • Geometry: Copying and structuring planar geometric figures.

  • Global Reach: Supporting nearly 100 languages.

DeepSeek-AI has publicly released the code and weights, paving the way for AI assistants that can "see" through massive libraries of data without breaking a sweat.

Would you like me to dive deeper into how DeepEncoder handles multilingual documents or its performance on specific STEM tasks like chemical formula parsing?