DeepSeek-OCR
DeepSeek's 3B MIT-licensed OCR model that packs page text into few vision tokens: about 97% decoding precision at under 10x compression.
In short. DeepSeek-OCR reads a picture of a page and writes out its text. The research idea behind it is that a page image can hold the same text in far fewer tokens than the words themselves, which could make very long inputs cheaper. It is small, MIT licensed and runs on one GPU.
Summary
DeepSeek released DeepSeek-OCR on 20 October 2025. It pairs a vision encoder called DeepEncoder with a DeepSeek3B-MoE-A570M decoder, a mixture-of-experts language model with about 3 billion parameters of which 570 million are active per token. The checkpoint holds about 3.34 billion parameters in BF16, at 6.67 GB, and the weights use the MIT license.
The paper frames the model as an initial investigation of compressing long contexts through optical 2D mapping. When the number of text tokens is within 10 times the number of vision tokens, it reports about 97% decoding precision, and at 20 times compression about 60%. On the OmniDocBench document benchmark it says the model beats GOT-OCR 2.0 using only 100 vision tokens per page, and beats MinerU 2.0, which averages more than 6,000 tokens per page, with fewer than 800. DeepSeek also says a single A100-40G can produce more than 200,000 pages of training data per day. These figures are from DeepSeek's own report.
Several resolution modes are supported: Tiny (512 x 512, 64 vision tokens), Small (640 x 640, 100 tokens), Base (1024 x 1024, 256 tokens), Large (1280 x 1280, 400 tokens) and a dynamic Gundam mode that tiles larger pages. The README documents prompts for Markdown conversion, free OCR, figure parsing and locating text, and reports about 2,500 tokens per second for concurrent PDF processing on an A100-40G. vLLM has supported the model upstream since 23 October 2025. DeepSeek followed it with DeepSeek-OCR 2 in January 2026 under the Apache 2.0 license.
Good for
- Document and PDF digitization on a single GPU with a clear MIT license
- Choosing a speed and accuracy trade-off through the resolution modes
- Studying optical context compression as a way to shorten long inputs
- Building pipelines on vLLM, which supports the model upstream
Watch out for
- Marked legacy because DeepSeek-OCR 2 (January 2026) is newer, uses a different encoder and is under Apache 2.0, even though this model still had about 1.9 million downloads in the 30 days to 11 November 2026, so compare the two on your documents
- Compression results describe OCR precision on specific test sets and do not guarantee the same accuracy on your pages
- The documented environment is CUDA 11.8 with PyTorch 2.6.0, and the model needs trust_remote_code
- Benchmark claims come from DeepSeek's own paper
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the OCR variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| OCR BF16 | 3.34B 570M active | 7.68 GB | 3.84 GB | 2.3 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is DeepSeek-OCR?
- DeepSeek-OCR reads a picture of a page and writes out its text. The research idea behind it is that a page image can hold the same text in far fewer tokens than the words themselves, which could make very long inputs cheaper. It is small, MIT licensed and runs on one GPU.
- Can I use DeepSeek-OCR commercially?
- Yes. DeepSeek-OCR is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is DeepSeek-OCR open source?
- DeepSeek-OCR is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run DeepSeek-OCR on my own computer?
- Yes. The smallest variant, OCR, needs about 2.3 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was DeepSeek-OCR released?
- DeepSeek-OCR was released by DeepSeek on October 20, 2025.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- DeepSeek-OCR model card (Hugging Face) huggingface.co
- DeepSeek-OCR: Contexts Optical Compression (technical report) arxiv.org
- DeepSeek-OCR repository, release notes and usage modes (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone | |
| Unlimited-OCR Baidu | 3.34B | – | MIT (Permissive) | 2.3 GB Laptop or phone | |
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server |