Post-90s Generation Takes the Center Stage: Yao Shunyu, Luo Fuli, Liu Dayiheng, Chen Yusen, Zhou Chang and Their Peers Take Over the Domestic Large Model Industry
In September 2026, domestic large model companies intensively released their latest performance reports.
On September 21, one day before the opening of the Yunqi Conference, Alibaba confirmed that Liu Dayiheng, born in 1993, serves as the head of the TokenFoundry Qwen large language model project under the ATH Business Group, and his speaking order on the main forum is only after Joseph Tsai, Chairman of Alibaba Group, and Wu Yongming, CEO of Alibaba Group.
On September 22, at the same conference, Chen Yusen, born in 1992, made his first public appearance as Vice President of Alibaba Group and CEO of Qwen Office.
Previously, he transformed from the founder of Chaitin Tech to CEO of DingTalk, then to the head of the Qwen Office Business Unit, with his job level promoted from P10 to P11 (M6).
On the same day, Xiaomi also released the MiMo-V2.6 series, while the internal promotion list was circulated: Luo Fuli, born in 1995, was promoted to level 22, the highest point of Xiaomi's job level system. The MiMo-V2.6-Pro she is responsible for scored 46 points in the Artificial Analysis Comprehensive Intelligence Index, ranking first among all global open-source weight models on this list.
On the other side, Zhou Chang, born in 1990, left Alibaba and joined ByteDance, where he is in charge of Seedream, Seedance and the world model.
Two generations appear alternately on the same organizational chart. On one side are technical veterans who took ten or twenty years to climb to the top position, and on the other side are the post-90s generation who compressed their resumes to five or even ten years.
In this technical cycle where "Transformer (the underlying architecture of large AI models) has been around for less than a decade and ChatGPT's explosion for less than five years", the variable of "seniority-based ranking" that has operated for decades seems to have lost its binding force.
But what the post-90s generation has really taken over is not only positions, but also unresolved technical debts, commercialization pressure and catch-up risks. The handover has already happened, and the exam has only just begun.
01
Liu Dayiheng and Chen Yusen: One Path Leads to the Technical Base, the Other Leads to Commercial Business
Placing Liu Dayiheng and Chen Yusen at the two consecutive days of the same Yunqi Conference clearly shows two completely different paths for post-90s executives in charge of large models.
Liu Dayiheng was admitted to the School of Computer Science of Sichuan University in 2012, was recommended for direct master-doctoral study with the first comprehensive ranking in 2015, and obtained his doctorate in 2020.
He first came into the public view when he published a paper on ACL in 2019, becoming the first in-service postgraduate student of Sichuan University to publish achievements on ACL.
In the same year, he participated in Google's Natural Questions global long-term public competition. His competitors included Google AI and IBM Research. The team he belonged to won the first place in the world at the time of submission and kept the leading position for three months.
In 2020, he became the only selected candidate in Southwest China in the first tier of Huawei's "Genius Youth Plan", and joined Alibaba Damo Academy in January 2021.
On March 9, 2026, he was responsible for the pre-training of Qwen, and concurrently managed the post-training and coding teams, reporting to Zhou Jingren, CTO of Alibaba Cloud, who was the interim head of Qwen. This appointment had been circulating inside Alibaba for half a year before it was officially announced to the public in September.
In the technical roadmap he announced, Recursive Self-Improvement (RSI) is the most radical approach: let the model discover deficiencies from real tasks, construct training data, and then evaluate and correct the results.
Chen Yusen took the position of head of Qwen Office through a path that started from technology and then moved to product development.
Born in 1992, he graduated from the Computer Science major of Qiushi Honors College, Chu Kochen Honors College of Zhejiang University, co-founded Chaitin Tech, an enterprise-level cybersecurity company, and was selected into the Forbes China 30 Under 30 list in 2017. In 2019, Chaitin Tech was fully acquired by Alibaba Cloud, and he joined Alibaba accordingly; later he left to start a business, and returned to Alibaba Cloud after the project was shut down.
On June 11 this year, Chen Yusen took over as CEO of DingTalk, becoming the youngest business unit CEO in Alibaba's history; in July, three originally scattered Agent product lines were integrated into the "Qwen Office" business unit, and he concurrently served as CEO.
Both of them are successors of star executives. Liu Dayiheng took over the vacant position of head of Qwen technical base left by Lin Junyang, while Chen Yusen took over the baton from Wuzhao (Chen Hang), who joined Alibaba in 2010 and was the first leader of DingTalk.
But the propositions they undertake are completely different.
What Liu Dayiheng needs to solve is the problem of the capability upper limit of the basic large model. The core challenge of base R&D is to continuously ease the contradiction between the model's capability ceiling and cost.
Chen Yusen's task focuses on the commercial implementation and integration of large models. He needs to embed large model capabilities into the original DingTalk workflow, find value points that enterprise customers are willing to pay for, and complete the transformation from technical capabilities to large-scale business revenue.
One path leads to the technical base, the other leads to commercial business. These two lines together form two sides of Alibaba's large model system.
02
Three Solid Performance Results: Open Source, Multimodal and Intelligent Agent
Luo Fuli, Zhou Chang and Yao Shunyiu are respectively in charge of the three directions of open-source large models, multimodal generation and AI Agent. Their resumes prove that in an industry with rapidly changing technical paradigms, one sufficiently key achievement is more convincing than years of step-by-step promotion.
Luo Fuli's story most directly reflects the promotion speed in the large model industry.
In November 2025, she officially joined Xiaomi and took full charge of the MiMo large model business. About 10 months later, in September 2026, she was promoted to level 22 of Xiaomi, the highest point of Xiaomi's job level system. There are 10 grades from level 13 to level 22 in Xiaomi, and vice presidents correspond to level 21 to level 22. Lei Jun himself is not included in this system.
This promotion is not only based on rumors and titles.
After graduation, Luo Fuli joined Alibaba Damo Academy and led the development of the multilingual pre-trained model VECO; in 2022, she joined High-Flyer, then joined DeepSeek as a deep learning researcher, and became one of the core developers of DeepSeek-V2, the MoE large model.
After joining Xiaomi, her performance results are more straightforward: MiMo-V2-Pro has a total parameter of more than 1 trillion, ranking fifth in the global ranking by brand on the Artificial Analysis list, surpassing xAI's Grok; MiMo-V2.6-Pro scored 46 points, ranking first among global open-source weight models.
Zhou Chang's story is more like a cross-company capability transfer.
In July 2017, he joined Alibaba Damo Academy through campus recruitment, with the flower name "Zhong Huang". In his first few years at Alibaba, he worked on e-commerce recommendation algorithms, which had nothing to do with large models.
Later, he led the model architecture design of "Tongyi-M6" and "Tongyi Qianwen". Qwen2, released on the eve of his resignation, topped the Hugging Face open source list within two hours, surpassing Llama 3-70B and many domestic closed-source models at that time.
Later, Alibaba filed a labor dispute arbitration on the grounds that he violated the non-compete agreement. This dispute once made the "non-compete wall" a public topic in the industry.
In the second half of 2024, Zhou Chang quietly joined ByteDance's Seed team at job level 4-2. Less than a year after joining, he took charge of the Seedream, Seedance and world model teams, and later his management scope further extended to the field of embodied intelligence.
What really brought him into the public view was Seedance 2.0 in February 2026. The product entered internal beta on February 7 and was officially released on February 12; during that period, A-share short drama concept stocks once rose by the limit collectively.
With a multimodal product, Zhou Chang opened a breakthrough for ByteDance's catch-up narrative.
Yao Shunyiu first defined the key methods of the Agent era in the academic stage, then joined top AI companies to participate in product implementation, and finally joined Tencent, in charge of both the model and infrastructure.
After graduating from Princeton University, Yao Shunyiu joined OpenAI in 2024 and participated in the development of AI Agent and task execution systems. In December 2025, Tencent upgraded its large model R&D architecture, and Yao Shunyiu became the Chief AI Scientist of the CEO/President's Office, concurrently serving as the head of the AI Infra Department and the Large Language Model Department.
AI Infra determines how large models are trained, how they are inferred, and how to improve efficiency; the Large Language Model Department determines the capability upper limit and iteration direction of Hunyuan. Tencent handed over these two parts to Yao Shunyiu at the same time, which means he is responsible for both the cutting-edge model capabilities and the engineering base that supports the model capabilities.
After joining Tencent, Yao Shunyiu completed a counterattack with Hy3. This medium-sized model with a total parameter of only 300B and an activation parameter of 21B outperformed GLM 5.1 in multiple capabilities.
After Hy3 went online, the number of users who actively chose Hy3 on Tencent's WorkBuddy increased by 6 times, and the average daily text processing volume increased by 20 times compared with the preview version.
The common point of Luo Fuli, Zhou Chang and Yao Shunyiu is that they achieved results first in a new paradigm, and then were granted greater authority by the organization.
03
The Handover Is Completed, and the Major Exam Has Just Begun
The phrase "post-90s taking over" has been repeatedly mentioned in the 2026 large model market, but it is not a finished result yet.
What has been completed is the position handover: Liu Dayiheng is in charge of the Qwen model, Chen Yusen is in charge of Qwen Office, Luo Fuli is in charge of MiMo, Zhou Chang is in charge of ByteDance's multimodal business, and Yao Shunyiu is in charge of Tencent's large language model and AI Infra.
What has not been completed is capability verification.
What Liu Dayiheng faces is the capability upper limit and cost constraint of the basic model. The difficulty of base R&D is not only to continue to expand the parameters, but to continuously make breakthroughs in long text, reasoning, code and agent capabilities, while controlling the huge computing power and data costs required for pre-training and post-training.
What Chen Yusen faces is commercial implementation and organizational integration. Qwen Office needs to package the scattered Agent capabilities, DingTalk ecosystem and Qwen base capabilities into products that users can perceive.
The real difficulty is not to demonstrate an Agent that can converse and call tools, but to let it enter real workflows such as meetings, documents, approvals and collaboration, solve the problem of "amazing in demonstration but difficult to use in daily scenarios", and find value points that enterprise customers are willing to pay continuously.
What Luo Fuli faces is whether the open-source achievements can be transformed into a sustainable paradigm.
She fully demonstrated the whole process of reinforcement learning training and made the cost public: the post-training of MiMo-V2.6-Pro took 6 days, used about 750,000 real task tracks, with a total training cost of 3.47 million US dollars, equivalent to about 4 million RMB per day.
This figure sparked discussions. Li Wangminghui, associate professor of the College of Electronic and Information Engineering of Tongji University, told the media that the key is not "low cost", but that it turns reinforcement learning post-training into a measurable service capability; but this depends on whether the marginal cost of environment construction, tool execution and result verification can be diluted.
If these heterogeneous environments can be engineered, standardized and reused, it may become a new service pricing paradigm; otherwise, 3.47 million US dollars is just a beautiful benchmark case.
What Zhou Chang and ByteDance face is how the catch-up party can make up for the lessons of basic capabilities.
Seedance 2.0 made the multimodal business led by Zhou Chang well-known, but this does not mean that the catch-up has ended. After video generation, there are a series of difficult problems such as world model, real-time interaction, cost control and large-scale commercial use.
What Yao Shunyiu faces is the test of capabilities extending to the organizational system. Tencent is not short of scenarios: social networking, content, meetings, documents, cloud and enterprise services, but the scenarios themselves will not automatically turn into capabilities. It is necessary to transform them into a model-understandable, executable and iterable system through tool access, data backflow, real evaluation and infrastructure construction.
This is exactly the other side of the post-90s takeover: they have entered a faster decision-making cycle, a larger resource pool, and a higher risk exposure. Positions can be handed over in one organizational adjustment, but capabilities must be verified through rounds of models, products, revenue and costs.
This article is from the WeChat official account "Tech Planet" (ID: tech618), author: Ren Xueyun, published with authorization from 36Kr.