TMTPOST -- Chinese artificial intelligence (AI) star DeepSeek appears to jump into the AI agent race, expanding its disruptive challenges to U.S. competitors including OpenAI. 

AI Generated Image

DeepSeek is developing an AI model with more advanced AI agent features, and its founder Liang Wenfeng is pushing his team to launch the new system in the fourth quarter of the year, Bloomberg cited anonymous sources on Thursday.  

AI agent is a kind of digital assistant that can complete tasks ranging for place orders online to coding without a lot of human oversight. The DeepSeek model is reportedly designed to carry out multi-step actions on a person’s behalf with minimal direction from the user. The system is also meant to learn and improve based on its prior actions.

The report suggested the model that DeepSeek is working on is a successor to DeepSeek-R1, a reasoning models that made the company a sudden rise in the AI race and shaked the industry.  DeepSeek claimed its performance comparable to leading offerings like OpenAI’s o1 at a fraction of the cost. The model is up to 50 times cheaper to run than many U.S. AI models. 

The free DeepSeek chatbot service application powered by the R1 model stunned the West and went viral overseas in late January.  It jumped to the No.1 spot in app stores on January 27, dethroning OpenAI’s ChatGPT as the most downloaded free app in U.S. on Apple Inc.'s App Store.  The app was downloaded 16 million times in its first 18 days, surpassing the 9 million downloads recorded by OpenAI's ChatGPT app in the same timeframe, data from Sensor Tower shows.

What also shocked Silicon Valley is that it took just $5.58 million for DeepSeek to train its V3 large language model (LLM) released in December. The Hangzhou-based company claimed it used 2,048 Nvidia H800 chips, a downgraded version of Nvidia’s H100 chips designed to comply with U.S. export restrictions.

The report on Thursday didn’t provide more details about DeepSeek’s upcoming mode. But the startup last month indicate it is foraying into true AI agents. It called its DeepSeek-V3.1 model “our first step toward the agent era.” 

DeepSeek’s domestic peers are making progress in tools to create AI agents these months. Alibaba Group’s agent framework Qwen-Agent was open-sourced in March. TikTok’s owner ByteDance in July open-sourced its agent development platform Coze Studio. Both of agent frameworks have alreday amassed more than 10,000 stars on GitHub, suggesting the high popularity. Tencent Holdings on Tuesday open-sourced its Youtu-Agent agentic framework on Tuesday.