DeepSeek-OCR 2

DeepSeek's 3.4B Apache 2.0 document OCR model, whose new encoder reorders visual tokens by meaning instead of reading in scan order.

In short. DeepSeek-OCR 2 reads document images and writes out their text or Markdown. Its new image encoder follows the logical order of a page, as a person would, instead of scanning from the top-left corner to the bottom-right. It is small enough for a single GPU, and the Apache 2.0 license allows commercial use.

Summary

DeepSeek released DeepSeek-OCR 2 on 27 January 2026 as the successor to DeepSeek-OCR. It has about 3.39 billion parameters and the BF16 checkpoint is about 6.8 GB. The weights use the Apache 2.0 license, while the first DeepSeek-OCR used MIT. Like its predecessor it has a 12-layer mixture-of-experts decoder, with 64 routed experts, 6 chosen per token and 2 shared, according to its configuration file.

The change is the encoder. Conventional vision-language models feed image patches to the language model in a fixed raster order, from top-left to bottom-right. DeepSeek's report introduces DeepEncoder V2, which reorders the visual tokens according to the meaning of the image before the language model reads them, an idea the authors compare to how people follow the logical structure of a page, and it describes the work as an investigation of this new design rather than a finished product. Resolution is dynamic: up to six 768 x 768 crops plus one 1024 x 1024 view give between 256 and 1,120 visual tokens per page.

The usage notes document two prompts, "Convert the document to markdown." and "Free OCR.", and offer vLLM and Transformers code. The tested environment is CUDA 11.8 with PyTorch 2.6.0 and FlashAttention. DeepSeek says PDF throughput in vLLM is on par with DeepSeek-OCR. The model card gives no benchmark table, so judge accuracy on your own documents and compare with GLM-OCR, PaddleOCR-VL and Unlimited-OCR, which publish their own benchmark results.

Good for

  • Converting scans, PDFs and screenshots with complex layouts into Markdown on a single GPU
  • Commercial document pipelines that need a clear Apache 2.0 license
  • Research into learned visual-token ordering and encoder design
  • Batch jobs served through vLLM

Watch out for

  • The model card and README publish no benchmark table, so measure accuracy on your own documents
  • The documented setup is old and specific, with CUDA 11.8, PyTorch 2.6.0, vLLM 0.8.5 and FlashAttention
  • It uses custom model code, so you must enable trust_remote_code
  • The license differs from the first DeepSeek-OCR, which is MIT, so check which one you deploy

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the OCR 2 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 2 BF16 3.39B 7.8 GB 3.9 GB 2.34 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is DeepSeek-OCR 2?
DeepSeek-OCR 2 reads document images and writes out their text or Markdown. Its new image encoder follows the logical order of a page, as a person would, instead of scanning from the top-left corner to the bottom-right. It is small enough for a single GPU, and the Apache 2.0 license allows commercial use.
Can I use DeepSeek-OCR 2 commercially?
Yes. DeepSeek-OCR 2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is DeepSeek-OCR 2 open source?
DeepSeek-OCR 2 is released under Apache 2.0, 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 2 on my own computer?
Yes. The smallest variant, OCR 2, needs about 2.34 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was DeepSeek-OCR 2 released?
DeepSeek-OCR 2 was released by DeepSeek on January 27, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
DeepSeek-OCR DeepSeek 3.34B 570M active – MIT (Permissive) 2.3 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
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone