DeepSeek-OCR

license

MIT(OSI

weights

open(downloadable from Hugging Face without a gate)

data

closed(the training corpus is neither published nor described)

code

closed(inference code is published

DeepSeek-OCR's code and weights are both released under MIT and download from Hugging Face without a gate, but DeepSeek does not publish or describe the training corpus or the pipeline used to build the model, only the inference code. That makes this an open-weights release rather than a fully open one.

• https://github.com/deepseek-ai/DeepSeek-OCR recorded 2026-09-16

  Repository page for deepseek-ai/DeepSeek-OCR: MIT, public and unarchived, 23,890 stars, carrying inference code. Establishes the licence, the published code and the absence of a training pipeline.

• https://huggingface.co/api/models/deepseek-ai/DeepSeek-OCR recorded 2026-09-16

  Hugging Face model API record: license mit, 2,386,735 downloads in the trailing 30 days, weights downloadable without a gate, and no dataset declared. Establishes the open weights and the absent corpus.

DeepSeek-OCR's adoption is measured by Hugging Face downloads over the trailing month, since Hugging Face is the declared distribution channel for the weights. Docling and MarkItDown, both distributed via PyPI, are measured with higher counts on the same reading.

• https://huggingface.co/api/models/deepseek-ai/DeepSeek-OCR recorded 2026-09-16

  Hugging Face model API record for deepseek-ai/DeepSeek-OCR: 2,386,735 downloads in the trailing 30 days, 3,380 likes, license mit.

DeepSeek-OCR converts a page image into text carried in vision tokens, a technique the release calls optical context compression. Neither the repository nor the model card says whether tables or reading order survive that conversion, so no capability score is assigned rather than assumed from silence.

• https://huggingface.co/api/models/deepseek-ai/DeepSeek-OCR recorded 2026-09-16

  The model's Hugging Face record: what is published, the declared licence, and - the point of citing it on an abstention - no statement about what document structure the output preserves.