Deepseek AI’s developers from China have introduced a new model that significantly enhances document processing efficiency. The latest version utilizes vision-text compression to convert text and documents into images. This process results in a dramatic reduction in token usage, with savings ranging from seven to 20 times, all while preserving high accuracy.
Since its debut in early 2025, Deepseek has impressed with its ability to match the capabilities of top AI models like OpenAI’s ChatGPT and Google’s Gemini. The AI achieves this efficiency despite requiring fewer resources for development. The latest innovation, Deepseek-OCR, represents a leap forward by minimizing token overhead, allowing the model to process large amounts of text more efficiently.
The developer noted that Deepseek-OCR’s vision-text compression technique enables significant token reduction across different historical contexts. They believe this approach offers a promising solution for managing long-context calculations in AI systems.
The new model integrates two key components: the DeepEncoder and DeepSeek3B-MoE-A570M decoder. The DeepEncoder transforms large volumes of text into high-resolution images. The decoder then analyzes these images, understanding the context with fewer tokens than if the text were directly processed.
This method excels at interpreting visual data like graphs and tables, which makes it valuable in fields such as finance, science, and medicine. The developers highlight its potential for improving data analysis in these sectors.
In performance tests, Deepseek-OCR demonstrated a 97% accuracy rate while reducing token usage by less than a factor of 10. When the token reduction reached 20 times, accuracy dropped to 60%. Though this suggests diminishing returns, even a small compression ratio could drastically lower the cost of running AI models.
The new model also has potential for generating training data for future AI systems. However, the introduction of even minor errors could undermine the model’s reliability, making this application challenging.
