Lin Junyang published his first long article after leaving Qianwen: AI is moving from "reasoning thinking" to "agent thinking"

Image source: Jiemian Image Library, Fan Jianlei

On March 26th, Lin Junyang, the former technical leader of Alibaba's Qianwen large model, published his first long post after leaving the company on a social platform. In the post, he didn't disclose any information related to his future career plans. Instead, he shared his technical insights, focusing on the development of large model technology and the prediction of the next stage of AI, which has attracted wide attention in the industry.

In this article titled From "Reasoning" Thinking to "Agentic" Thinking, Lin Junyang stated that the past two years have reshaped the industry's evaluation methods and core expectations for large models.

He pointed out that in the first half of 2025, the industry's focus mainly remained on "reasoning thinking" itself. That is, through technologies such as reinforcement learning, the model generates a rigorous reasoning chain through internal logical deduction after receiving a given input, thereby improving the accuracy of the answers. OpenAI's o1 and DeepSeek - R1 are typical representatives of this paradigm. "In the first half of 2025, the industry was thinking about how to make the model 'think a little longer'. Now it's time to ask about the next step."

He judged that "agentic thinking" is the core direction of the next stage of AI. The model is no longer limited to internal static reasoning. It thinks for action and continuously updates its plans according to the feedback from the world during the interaction with the environment.

Lin Junyang also shared the practical experience of the Qianwen team. He mentioned that at the beginning of 2025, the Qianwen team had a big ambition: to create a unified system that combines the thinking mode and the instruction mode. This direction is correct. However, he also said that the current combination of the two modes in practice is not very successful. The core problem is not the model compatibility, but the significant differences in the data distribution and behavioral goals of the two modes: the reasoning model relies on logically rigorous and verifiable data, while the agent model requires complete interaction trajectory data. "A truly successful combination requires a smooth spectrum of reasoning intensity."

He believes that a longer reasoning chain does not necessarily mean a smarter model. In many cases, a longer reasoning chain actually indicates that the model is wasting computing power.

He mentioned that the Qianwen team realized that the industry is moving from the era of training models to the era of training agents, whose defining feature is the closed - loop interaction with the real world. The core question of agentic thinking has changed from "can the model think long enough" to "can the model think in a way that can support effective actions".

He predicts that agentic thinking will become the mainstream thinking form. Of course, the model architecture and training data are still important, but the environment design, rollout infrastructure, the robustness of the evaluator, and how to coordinate multiple agents have all entered the core circle.

On March 4th, Lin Junyang publicly announced on social media that he had left Qianwen. On the same day, Yu Bowen, the post - training leader of Qwen, and Li Kaixin, a core contributor to Qwen 3.5/VL/Coder, also revealed their departure news.

On March 5th, Wu Yongming, the CEO of Alibaba Group, responded in an internal email that the company had decided to approve Lin Junyang's resignation and thanked him for his past contributions in his position. Zhou Jingren, the CTO of Alibaba Cloud and the head of Tongyi Lab, will continue to lead the Tongyi Lab to advance the follow - up work. At the same time, the company will establish a basic model support team to coordinate the group's resources to support the construction of basic models.

Lin Junyang was born in 1993. He was once the youngest P10 - level technical expert at Alibaba. He graduated from the School of Foreign Languages of Peking University with a master's degree. He joined Alibaba's DAMO Academy as a senior algorithm engineer in 2019 and gradually participated in the R & D of core models such as M6 and OFA. After the establishment of Tongyi Lab at the end of 2022, he was appointed as the technical leader of the Qianwen large model and led the team to create the Qwen series of open - source models. As of his departure, the global download volume of the Qwen series had exceeded 600 million times, and there were more than 200,000 derivative models.

This article is from Jiemian News, author: Shen Xiaoge. Republished by 36Kr with permission.