DeepSeek, the Chinese AI giant, is back in the headlines as the company has unveiled its latest V3.1 model. The company says that this is the first step toward the “agent era.” The release introduces hybrid reasoning modes, stronger agent abilities, and a range of API upgrades. It will also compete against OpenAI’s latest GPT-5 AI model. So, in this article, I have compiled 5 things that you should know about the newly arrived V3.1.
1. Hybrid Inference with Two Modes
One of the highlighted features of the DeepSeek V3.1 is its hybrid inference system. The model now supports both “Think” and “Non-Think” modes, allowing you to choose between deeper reasoning tasks and faster, lighter responses. The company has also introduced a new DeepThink button, which makes toggling between the two modes straightforward.
2. Tools and Agent Upgrades
As indicated already, with this launch, the company is pushing harder into the agent space with big improvements in multi-step reasoning and task execution. The model also shows better performance on benchmarks like SWE and Terminal-Bench, which hints at its ability to handle complex search and agent-driven workflows. Apart from that, DeepSeek also claims major improvements in “thinking efficiency,” which should benefit users running heavier reasoning tasks.
3. Faster and Smarter Than Its Predecessor
According to the company, the new V3.1 is a massive upgrade over the previous R1-0528. The model processes answers faster in Think mode, while also benefiting from post-training upgrades that make it better at handling tool use and multi-step agent tasks.
4. API Improvements
The new release also comes with multiple API improvements for developers. DeepSeek has split endpoints into deepseek-chat for non-thinking mode and deepseek-reasoner for thinking mode, both supporting 128K context. The model now works with Anthropic’s API format, which makes it easier for developers to adopt, and strict function calling is supported in beta.
5. Model Training and Pricing Updates
Under the hood, the new V3.1 model is built on the V3 base with 840 billion tokens of continued pretraining, which extends its long-context capabilities. Besides, the tokenizer and chat template have also been updated to better handle different inputs. On the pricing side, the current structure remains in place until September 5, 2025, after which new rates will kick in and off-peak discounts will end.
That’s all about DeepSeek V3.1 for now. But if you want to stay updated with all the latest tech and AI updates, then join us on WhatsApp, where we, along with the news, also share our reviews, analysis, and more.
