DeepSeek-OCR

What it is

DeepSeek-OCR is an open-source optical character recognition model released on 20 October 2025 under the MIT License, with a paper posted to arXiv on 21 October 2025.^[1][2] The Hugging Face model card lists it at 3B parameters in BF16.^[3] What makes it more than another OCR system is the question it is built to probe: the paper frames its subject as "contexts optical compression", meaning the rendering of long text as 2D images so that a model can read the text back out of far fewer vision tokens.^[2]

In other words, document recognition is used here as a measurable test case for a broader idea — that a page of text can be stored and processed as a picture, and that the picture can be much smaller than the equivalent string of text tokens would be.

How it works

The architecture has two parts: a DeepEncoder vision encoder and a DeepSeek3B-MoE-A570M decoder, a mixture-of-experts design of about 3B parameters of which 570M are activated.^[2] The encoder turns a page image into a small set of vision tokens; the decoder writes the text back out.

Users choose how many vision tokens a page is worth. The native-resolution modes are Tiny at 512×512 (64 vision tokens), Small at 640×640 (100), Base at 1024×1024 (256) and Large at 1280×1280 (400), plus a dynamic "Gundam" mode that tiles images.^[1] In the model card's usage code, the Gundam preset uses base_size 1024, image_size 640 and crop_mode enabled, and it is the only preset there with cropping turned on.^[3]

The number that matters is the compression ratio: how many text tokens the model recovers per vision token. A page read at 10× compression carries ten times as much text as its vision token count would suggest, and the model has to reconstruct the rest from the image.

Results

The paper reports how decoding precision falls as compression rises. When text tokens are within 10 times the number of vision tokens, OCR decoding precision is about 97%; at 20× compression it is about 60%.^[2] That curve is the substance of the argument: visual tokens are not free, but a great deal of text can be carried by relatively few of them before accuracy degrades badly.

On OmniDocBench the model beat GOT-OCR2.0, which uses 256 tokens per page, while using only 100 vision tokens; it also beat MinerU2.0, which uses more than 6,000 tokens per page, while using fewer than 800.^[2] The model card reports an olmOCR-bench overall score of 75.7 and an overall score of 73.99 on the PaddlePaddle Real5-OmniDocBench set.^[3]

Beyond benchmarks, the paper describes a production use: generating training data at over 200,000 pages per day on a single A100-40G GPU.^[2] That is a scale at which document images become a practical source of text for training other models, rather than a curiosity.

Release and what followed

Besides the MIT-licensed release on GitHub,^[1] the model was supported in upstream vLLM from 23 October 2025,^[1] which let it be served with the same inference stack many teams already use. A successor, DeepSeek-OCR 2, was presented on 27 January 2026.^[1]

Read together, the pieces describe a model whose output is text but whose lasting contribution is a measurement: how much text fits into a vision token, and at what point the reconstruction stops being faithful. Treating OCR as an optical compression problem gives a clear way to answer that, and a way to trade tokens against fidelity when applying the model.