DeepSeek-V3.1 Officially Released, Dubbed as First Step towards Agent Era

Xinhua Tech News August 21, DeepSeek officially released its new-generation large language modelDeepSeek-V3.1, dubbed as "First Step towards Agent Era".

DeepSeek-V3.1 This upgrade includes the following major changes:

Mixed Inference Framework: A model that supports both thinking mode and non-thinking mode simultaneously;

Highest Thinking Efficiency: Compared to DeepSeek-R1-0528, DeepSeek-V3.1-Think can provide answers in a shorter time frame;

Stronger Agent Capabilities: Through post-training optimization, the new model shows significant improvements in tool use and intelligent agent tasks.

As of now, the company's official App and web services have completed the upgrade, allowing users to switch between two modes using the new "Deep Thinking" button.

According to official data, the new model's "Deep Thinking" mode shows higher efficiency in handling complex problems. Compared to the previous generation (DeepSeek-R1-0528), V3.1 can reduce output token quantity by 20% to 50%, while maintaining performance, meaning users can obtain high-quality answers with shorter time and lower costs.

Commercial strategy-wise, DeepSeek announces that it will adjust the prices of its API interfaces. The new pricing system will take effect onBeijing Time September 6, 2025 at midnight, when all previous night-time discounts will also be cancelled.