News 250821

DeepSeek-V3.1 was released as a first step toward the agent era. At its core is hybrid reasoning that switches between two modes, Think and Non-Think, within a single model.

• DeepSeek-V3.1-Think has been improved to produce answers in less time than DeepSeek-R1-0528.
• Post-training has also strengthened performance in tool use and multi-step agent tasks.
• Users can switch between Think/Non-Think via the DeepThink button on the Chat screen.

The API has been updated as well. deepseek-chat connects to non-thinking mode and deepseek-reasoner to thinking mode, with both supporting a 128K context. In addition, support for Anthropic API format and beta support for Strict Function Calling have been added.

On the model side, V3.1 Base was released, with long-context extension carried out through continued pretraining on 840B tokens. The tokenizer and chat template have also been updated, and open-source weights as well as Base version weights are each provided.

A pricing policy change has also been announced. New pricing will take effect from 16:00 UTC on September 5, 2025, and off-peak discounts will end; existing pricing will remain in place until then.