The Hangzhou-based startup is building an AI model that’s designed to carry out multi-step actions on a person’s behalf with minimal direction from the user, said the people. The system is also meant to learn and improve based on its prior actions, the people said, speaking on condition of anonymity because the information is private.
DeepSeek joins a broader industry movement to create true AI agents, considered the next stage of evolution in artificial intelligence. Liang Wenfeng, DeepSeek’s founder, is pushing his team to unveil the new software in the final quarter of this year, the people said. The impending release comes as the industry awaits the release of a successor to its R1, whose emergence in January upended the global tech industry.
That seminal platform, which mimics the human process of reasoning, purportedly cost just several million dollars to build yet matched or surpassed OpenAI products in benchmark tests. The agentic AI platform in the works underscores Liang’s determination to remain at the forefront of the highly competitive industry, though it’s unclear how or whether it can replicate the R1’s breakthroughs.
Since the R1, the Chinese upstart has only put out minor upgrades, even as rivals in the US and China have launched a flurry of new models. Local media have attributed the R2’s delay to Liang’s determination to get it right, even as he continues to run his lucrative High-Flyer Asset Management outfit. Others have speculated about various glitches in training or development.
DeepSeek did not respond to two separate emails seeking comment.
The company’s plan for a new agent-focused model, which hasn’t been previously reported, mirrors a broader shift in the tech industry.
OpenAI, Anthropic, and Microsoft have each launched their own takes on agent software in recent months to streamline personal and professional tasks. Manus AI, a Chinese-founded startup, also gained global attention with what it calls a general AI agent.
Rather than simply respond to a user with a few pithy lines of text, as many chatbots do, this newer crop of services is intended to field more complex tasks, ranging from researching a vacation to writing and debugging computer code. In a recent research report, Goldman Sachs said the next landmark advance in generative AI would be to boost business productivity through AI agents.
The goal for DeepSeek, and much of the industry, is to build increasingly autonomous AI systems that can initiate and execute complicated real-world action with little need for human intervention. To date, however, AI agents often require a fair amount of adult supervision.
Unlike its Chinese rivals, DeepSeek has maintained a fairly muted pace of development. That’s as rivals from Alibaba Group Holding to Tencent Holdings have kept up a frenetic pace of AI model rollouts and updates. Alibaba’s Qwen models have, in particular, gained a popular following.
