DeepSeek-OCR
Optical character recognition model using visual compression
DeepSeek-OCR is a specialised model for optical character recognition that compresses visual data into text tokens. Released in October 2025 by researchers including Haoran Wei, it uses a unique encoder-decoder design to handle high-resolution documents efficiently. The model aims to reduce the computational cost of processing long visual contexts while maintaining accuracy.
- Uses DeepEncoder for high visual compression ratios
- Decoder has 3 billion parameters with 570 million active
- Achieves 97% OCR precision at under 10x compression
- Processes 200,000+ pages per day on a single A100 GPU
- Outperforms GOT-OCR2.0 using fewer vision tokens
What it is
DeepSeek-OCR is a multimodal model designed for optical character recognition through visual context compression. It was introduced in October 2025 by a team including Haoran Wei, Yaofeng Sun and Yukun Li. The system combines a custom DeepEncoder with a decoder based on the DeepSeek3B-MoE architecture. Its primary function is to convert images of documents into markdown or text formats while significantly reducing the number of vision tokens required. This approach allows for efficient processing of high-resolution inputs without overwhelming the model's context window.
How it came about
The model was announced in late October 2025 via an arXiv preprint titled DeepSeek-OCR: Contexts Optical Compression. The work represents an initial investigation into compressing long contexts using optical two-dimensional mapping. Prior to its official release, the developers acknowledged contributions from projects such as Vary, GOT-OCR2.0 and MinerU. By late October 2025, the model gained support in the upstream vLLM library, facilitating faster inference for users.
Inside the model
The architecture consists of two main parts: the DeepEncoder and a decoder named DeepSeek3B-MoE-A570M. The encoder acts as the core engine, maintaining low activations even with high-resolution inputs to achieve high compression ratios. The decoder is a mixture of experts model with 3 billion total parameters, of which 570 million are active per token. This design ensures the number of vision tokens remains manageable. The model can operate in various modes, including Tiny, Small, Base and Large, depending on the base and image sizes selected during inference.
What it can do
Even at a twenty-fold compression ratio, accuracy remains around 60 per cent. On the OmniDocBench benchmark, it outperforms GOT-OCR2.0 by using only 100 vision tokens per page compared to 256. It also surpasses MinerU2.0 while using fewer than 800 vision tokens, whereas the latter uses over 6000 on average. In production settings, a single A100-40G GPU can generate training data for over 200,000 pages daily.
How to use it
The model weights and code are publicly available on Hugging Face under an open licence. Users can run inference using Hugging Face transformers on NVIDIA GPUs with CUDA support. It is also integrated into the vLLM framework for accelerated inference, supporting batched input processing. The developers provide guidance for installation and usage, including specific requirements for Python and flash attention libraries.
Why it matters
DeepSeek-OCR demonstrates potential for research into historical long-context compression and memory forgetting mechanisms in large language models. By proving that high compression ratios can retain reasonable accuracy, it offers a new pathway for handling visual data efficiently. Its ability to generate large-scale training data quickly may influence how future vision-language models are trained on document datasets.
Questions people ask about DeepSeek-OCR
What is the main purpose of DeepSeek-OCR?
It is designed for optical character recognition by compressing visual document data into a small number of text tokens. This allows for efficient processing of high-resolution images while maintaining high accuracy in text extraction.
How does the model architecture work?
It uses a two-part design with a DeepEncoder for visual compression and a DeepSeek3B-MoE decoder for text generation. The decoder is a mixture of experts model with 3 billion total parameters but only activates 570 million per token.
Is DeepSeek-OCR available for public use?
Yes, the model weights and code are openly available on Hugging Face. It supports inference via Hugging Face transformers and is also integrated into the vLLM library for accelerated performance on NVIDIA GPUs.
Good reading and sources
There is no English Wikipedia article on DeepSeek-OCR yet, so the facts on this page come from the sources listed above: huggingface.co/deepseek-ai/DeepSeek-OCR, arxiv.org/abs/2510.18234. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.