On August 21, DeepSeek Technology Co., Ltd. officially released its latest version of AI model - DeepSeek-V3.1. This upgrade marks a solid first step for the company towards the Agent era, providing users with a more powerful, efficient, and multifunctional artificial intelligence solution.

The release of DeepSeek-V3.1 brings multiple significant improvements. First, this version introduces a hybrid reasoning architecture, allowing a single model to support both thinking mode and non-thinking mode, offering users more flexible choices. In thinking mode, DeepSeek-V3.1-Think can provide answers in a shorter time compared to the previous DeepSeek-R1-0528 version, significantly improving thinking efficiency. Additionally, through Post-Training optimization, the new model has also seen a significant improvement in performance in tool usage and agent tasks.

To better serve users, the official app and web version models have been upgraded to DeepSeek-V3.1. Users can easily switch between thinking mode and non-thinking mode by clicking the "Deep Thinking" button. At the same time, the DeepSeek API has also been upgraded synchronously, where `deepseek-chat` corresponds to the non-thinking mode, and `deepseek-reasoner` corresponds to the thinking mode. The context capacity has been expanded to 128K. In addition, the API Beta interface has added support for `strict` mode Function Calling, ensuring that the output functions meet the schema definition.

DeepSeek-V3.1 also enhances support for the Anthropic API format, allowing users to easily integrate the model's capabilities into the Claude Code framework. This provides developers with greater flexibility and further expands the application scenarios of the model.

WeChat screenshot_20250821145433.png

In terms of tool calling and agent support, DeepSeek-V3.1 has made significant progress. In the programming agent evaluation, whether using an internal framework or performing complex task tests in a command-line terminal environment, DeepSeek-V3.1 shows noticeable improvements over the previous version. In the search agent evaluation, DeepSeek-V3.1 has achieved significant improvements in multiple indicators, especially in complex search tests requiring multi-step reasoning and interdisciplinary expert-level problems, where its performance is significantly ahead of the previous version.

Additionally, DeepSeek-V3.1 has also optimized thinking efficiency. After chain-of-thought compression training, V3.1-Think maintains the same average performance as R1-0528 while reducing the number of output tokens by 20%-50%. At the same time, in non-thinking mode, V3.1 can effectively control the output length, maintaining the same model performance as previous versions.

To further promote the open source and sharing of technology, the Base model of DeepSeek-V3.1 has been open-sourced on Huggingface and ModelScope platforms. The Base model has been re-trained and expanded based on V3, increasing the training volume by 840B tokens.

In terms of pricing, DeepSeek Technology Co., Ltd. announced that starting from 0:00 Beijing Time on September 6, 2025, it will adjust the pricing for the DeepSeek Open Platform API interface calls, implement the new pricing table, and cancel the night-time discount. Before September 6, all API services will still be billed according to the original pricing policy, and users can continue to enjoy current discounts. Meanwhile, the company has further expanded API service resources to better meet users' call demands.

Open source links:

https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base

https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Base

Official documentation:

https://api-docs.deepseek.com/zh-cn/guides/anthropic_api