At a recent summit hosted by Tsinghua University, industry experts declared a consensus: the era of AI as a mere chatbot is over. The new frontier is agentic AI—systems that can plan, execute, and learn from tasks autonomously.
This shift is not just philosophical. It is being driven by concrete technical innovations. Chinese researchers are moving away from brute-force scaling of model parameters toward what they call the "density law": extracting more intelligence with less compute and data. Sparse attention mechanisms, such as DeepSeek's NSA and Moonshot AI's MoBA, are key enablers, allowing models to focus on critical information rather than processing everything equally.
"The scaling law is not dead, but the marginal benefit of piling on more compute is diminishing," said Zhang Yaqin, founding dean of Tsinghua's Institute for AI Industry Research. The industry is now prioritizing efficiency and real-world utility.
Tencent, for example, has deployed its proprietary large model across over 900 internal scenarios, embedding AI deeply into its ecosystem. Baidu has restructured into separate foundational and applied model R&D departments. Founder Robin Li predicted that only a few foundational models will survive, but the application layer will see a proliferation of successful players.
The "hundred-model war" is winding down. Among the "Six Little Tigers" of Chinese AI, Baichuan Intelligence has pivoted to healthcare, while Lingyi Wanwu focuses on enterprise customization. The race is now about penetration into real-world scenarios and building industrial ecosystems.
Agentic AI, experts say, is the next evolution. Unlike chatbots that merely respond, agents can set goals, plan paths, and learn from feedback. They are autonomous, capable of generalization, and have long-term memory. "If a chatbot is a talking dictionary, an agent is a butler who can work independently," said Huang Jin, a researcher at the Chinese Academy of Sciences.
But challenges remain. Wei Kai, director of the AI Research Institute at the China Academy of Information and Communications Technology (CAICT), noted that agents still struggle with reliability, context memory, and long-horizon tasks. Large-scale deployment is still some way off.
Looking further ahead, Zhang Yaqin and others see AI expanding beyond the digital realm into the physical and biological worlds. In January, a Chinese embodied intelligence model achieved top scores in a global benchmark, signaling progress in giving robots the ability to understand and act in the physical world.
As China enters the 15th Five-Year Plan period, the government is pushing for AI integration across industries, culture, social governance, and livelihoods. The message from the summit is clear: the next phase of AI competition will be won not by the biggest model, but by the most useful one.
Why it matters
- Marks a strategic pivot from model scale to practical utility in China's AI industry
- Opens up new opportunities for application-layer innovation and enterprise adoption
- Signals a maturing market where differentiation comes from real-world integration, not just parameter count
What to watch next
- How quickly agentic AI systems overcome reliability and memory limitations for large-scale deployment
- Which companies successfully transition from foundational models to vertical applications
- The impact of density law and sparse attention on reducing compute costs and democratizing AI access
Sources
- 2026年中国AI发展趋势前瞻-清华大学 (Tsinghua University)
Confidence: 85%