This article first appeared on GuruFocus.
DeepSeek’s next move could reshape the AI landscapeagain. According to people close to the matter, the Hangzhou-based startup is preparing to launch a next-gen AI agent capable of executing multi-step tasks with minimal user input and improving over time through autonomous learning. The effort marks a major shift toward agentic AI, a field US leaders like OpenAI, Anthropic, and Microsoft (NASDAQ:MSFT) have already jumped into. DeepSeek’s founder, Liang Wenfeng, is pushing for a Q4 release, positioning the new system as a successor to R1, the January-launched platform that punched well above its weight and rivaled OpenAI in benchmark testsdespite costing just a few million to build.
Since R1’s breakout, however, DeepSeek has largely stayed quiet. While Chinese peers like Alibaba and Tencent ramped up with frequent model updatesAlibaba’s Qwen models in particular building strong tractionDeepSeek focused inward. The delay of R2 has sparked speculation: some cite technical setbacks, others point to Liang’s perfectionist streak. Still, the ambition remains clear. This new agent isn’t just another chatbotit’s being designed to handle real-world tasks like writing and debugging code, researching trips, and initiating workflows, all with limited oversight. That said, much of the industry acknowledges current AI agents still require a fair amount of “adult supervision.”
Whether DeepSeek’s agent can deliver remains the billion-dollar question. The company’s minimalist update cadence stands in contrast to its rivals’ rapid-fire model rollouts. But Liang, who also leads High-Flyer Asset Management, appears unfazed. His goal, insiders say, is not just another flashy demoit’s building something foundational. The agent’s successor failurecould determine whether DeepSeek regains its edge in a market now defined by blistering pace and relentless iteration. The world will be watching Q4.