Startup DeepSeek has unveiled its new model V3.1, which the company claims provides superior performance compared to the R1. The update includes specialized thinking modes that mark an important development step in AI technology.
Key Features of Model V3.1
DeepSeek claims that the new model V3.1 offers quicker responses to queries and improved agent capabilities. It supports a hybrid architecture of thinking and non-thinking, allowing users to switch between modes via a 'Deep Thinking' button. Compared to model R1, V3.1 shows better results on tests like SWE and Terminal-Bench, along with higher reasoning efficiency.
Pricing and Policy Structure
With the V3.1 update, DeepSeek also introduced a new pricing policy effective from September 6. Input API prices are set at $0.07 per million tokens for cache hits and $0.56 for misses, while output tokens cost $1.68 per million. These rates are significantly lower than competitors such as Gemini and OpenAI.
Expectations for Model R2 and Hardware Challenges
Despite the success of model R1, analysts are expecting DeepSeek to launch the successor R1, model R2, which was anticipated earlier this year. Founder Liang Wenfeng has expressed dissatisfaction with delays stemming from technical issues with Huawei processors. Reports suggest that DeepSeek is using Nvidia processors for training while relying on Huawei processors for inference.
Overall, with the new model V3.1, DeepSeek is taking steps to solidify its position in the AI market, though issues with hardware and expectations for model R2 remain key concerns for the company.
