DeepSeek releases DeepSeek-OCR, compressing text into images

Background

Language models process text as sequences of tokens, and long documents produce long sequences that are costly to handle. Optical character recognition, or OCR, faces a related problem: a large volume of text often arrives as page images, and the system has to turn those images back into text. DeepSeek's research asked whether the image itself could serve as a compact carrier for long text.

What happened

On 20 October 2025 DeepSeek released DeepSeek-OCR under the MIT License ^[1][2]. The model was built to explore what the company calls contexts optical compression: rendering long text as two-dimensional images, so that a model can read the content back from far fewer vision tokens ^[2]. The research paper was posted to arXiv on 21 October 2025 ^[2].

Key details

The system pairs a vision encoder named DeepEncoder with a decoder named DeepSeek3B-MoE-A570M. The decoder has about 3 billion parameters, of which about 570 million are activated, meaning used, for each step ^[2].

The central measurement concerns compression and accuracy. When the number of text tokens is within 10 times the number of vision tokens, OCR decoding precision is about 97 percent. At 20 times compression, precision is about 60 percent ^[2]. The result shows that the approach is accurate at moderate compression and degrades as compression increases.

On OmniDocBench, DeepSeek-OCR beat GOT-OCR2.0, which used 256 tokens per page, with only 100 vision tokens. It also beat MinerU2.0, which used more than 6,000 tokens per page, with fewer than 800 ^[2].

In production use, the model can generate training data at more than 200,000 pages per day on a single A100-40G GPU ^[2]. That throughput is the practical side of the approach: a model that reads documents with fewer tokens is cheaper to run at scale.

Impact

The release gave the research idea a public, openly licensed model that others could test. The paper's comparisons suggest that a document page can be represented with a few hundred vision tokens, where other systems needed many more ^[2]. The compression-accuracy curve also sets a clear limit: beyond moderate compression, precision falls sharply, so the method is a trade-off rather than a free saving ^[2].