On August 21st, DeepSeek Technology Co., Ltd. officially launched its latest AI model, DeepSeek-V3.1, marking a significant step forward in the company's intelligent agent technology. This upgraded model utilizes an innovative hybrid inference architecture, enabling for the first time the ability to freely switch between thinking and non-thinking modes, providing users with a more efficient and flexible intelligent experience.
The most notable change in the new model is its comprehensive performance improvements. The thinking mode, DeepSeek-V3.1-Think, is more responsive and significantly more efficient than its predecessor, version R1-0528. Post-training optimizations have also significantly improved the model's performance in tool usage and agent tasks. The official app and website have been updated simultaneously, allowing users to switch modes with a simple click of the "DeepThink" button. The API has also been upgraded to 128KB of context capacity, and strict mode function calling support has been added.
In specific applications, DeepSeek-V3.1 demonstrates its strong technical capabilities. Whether evaluating programming agents or testing search agents, the new model significantly outperforms previous models on complex tasks and multidisciplinary challenges. Particularly noteworthy is that after training with thought chain compression, V3.1-Think maintains comparable accuracy to R1-0528, even with output reduced by 20%-50%.
To promote technology sharing, DeepSeek has open-sourced the base version of this model on the Huggingface and Moda platforms. The company also announced that it will implement a new API pricing policy starting September 6th, but users will still enjoy the current discounts before the adjustment. With the launch of DeepSeek-V3.1, the technological boundaries of AI assistants have been further expanded, laying a solid foundation for the future development of intelligent agent technology.
