Chinese AI startup DeepSeek has introduced DeepSeek V3.1, a major update to its V3 model, which the company calls its first step into the “agent era.”
In a Weixin post announcing the launch, DeepSeek said the update introduces a “hybrid reasoning architecture: a model supports both thinking mode and non-thinking mode.” Users can toggle this feature through a new Deep Thinking button available on the app and web platform.
The V3.1 model is designed to handle tasks more efficiently than its predecessor. DeepSeek noted that, compared with the earlier R1-0528 model, “DeepSeek-V3.1-Think can provide answers in a shorter time.”
Beyond speed, the update enhances agent-style performance, allowing the model to carry out tool usage and programming tasks more effectively. In coding benchmarks such as SWE-bench Verified, the model outperformed DeepSeek’s R1-0528 with a score of 66.0%, compared to R1’s 44.6%.
Domestic chip optimization
One of the most significant technical changes? V3.1 is optimized to work with China’s upcoming semiconductor ecosystem.
The model now supports the UE8M0 FP8 precision format, allowing it to run faster and with less memory. This focus on homegrown hardware compatibility comes as Beijing pushes to reduce reliance on US technology amid ongoing export restrictions.
Open source and API expansion
DeepSeek announced that the base model of V3.1, retrained on 840 billion tokens, has been made available on Hugging Face and ModelScope, with both base and post-trained versions accessible to developers.
The API has also been upgraded, with support for longer 128K context windows and new function-calling capabilities. Developers can now integrate DeepSeek more easily, including with Anthropic API formats, broadening compatibility with existing frameworks.
The company warned that V3.1 introduces significant changes to its tokenizer and chat templates compared with earlier versions, urging developers to carefully review documentation before deploying it.
Pricing changes in September
Alongside the launch, DeepSeek introduced new pricing for its API services. The new structure will take effect on Sept. 6, 2025.
DeepSeek is a standout among AI competitors
DeepSeek gained global attention in January when its R1 model shook the industry with strong performance at lower costs. The company’s rapid progress has kept it in competition with Western players like OpenAI, even without access to the most advanced semiconductors.
DeepSeek V3.1 signals the startup’s move into the agent-driven future of AI.
In the race for AI dominance, explore how Deepseek’s models compare with those from competitors such as OpenAI and Alibaba.