The first batch of 70 AI digital "employees" developed using Deepseek launched on “Happy Futian,” the official WeChat account of the Futian District Government, Feb. 17. The personalized intelligent agents are tailored to integrate into the business processes of various departments, supporting a total of 240 scenarios.

The Futian District Government Large Model 2.0 is based on the DeepSeek R1 large model and addresses significant challenges, including high computational consumption, unstable responses, and a lack of specialization often seen in traditional government models.

These 70 digital "employees" cover 11 major categories spanning the entire spectrum of government services, such as document processing, public services, emergency management, and investment attraction. With the deployment of the large model 2.0, the time required for customization has been reduced from five days to just minutes.

The accuracy of document format corrections exceeds 95%, with review times shortened by 90% and error rates maintained below 5%. The accuracy of dispatching public demand requests has surged from 70% to 95%.

On Sunday, Shenzhen officially launched application services powered by DeepSeek-R1 for all government agencies across the city, becoming the first city in Guangdong Province to do so.

The integration of the DeepSeek-R1 model with the government cloud will significantly enhance services, such as providing 24/7 intelligent consulting for residents through the model’s natural language processing capabilities. It will also support data analysis and decision-making, efficiently handling large amounts of government data, among other improvements.

Released by Chinese AI company DeepSeek on Jan. 20, the DeepSeek-R1 model is a first-generation reasoning model developed through large-scale reinforcement learning. It is specifically designed to tackle complex reasoning tasks across various domains, including mathematics, coding, and language processing.

Currently, approximately 100 companies, including Tencent, have already integrated with the DeepSeek large model.