Byte Seed poaches key talent from Qianwen.

``` AI industry competition and convergence are also reflected in the movement of talent. On March 12, after Lin Junyang, the former head of technology for Alibaba Tongyi Laboratory's Qwen large model, resigned, the next destination of another core member of his team finally came to light. Industry sources say that Yu Bowen, the former head of post-training for Qwen, has officially joined ByteDance, serving as the head of post-training for the Vision Model and Multimodal Interaction Team in the Seed Team. People close to ByteDance confirmed the personnel change to Wallstreetcn. This personnel change comes at a time when the Alibaba Qwen team has just completed an organizational restructuring and several core technical talents have left intensively. It has drawn widespread industry attention to talent mobility and competition in technical routes in China's large model field. Yu Bowen’s academic and technical background are widely recognized as solid within the industry. Public information shows he graduated from Central South University with an undergraduate degree, then attended the Institute of Information Engineering, Chinese Academy of Sciences for graduate studies, and obtained a PhD from the University of Chinese Academy of Sciences in 2022. During his PhD, he focused on natural language processing and information extraction, published multiple papers at top international conferences like ACL and EMNLP, and creatively proposed transforming information extraction tasks into graph structure problems. This effectively solved entity overlap and nesting challenges in complex scenarios. Owing to his outstanding academic performance, he was awarded the President’s Award of the Chinese Academy of Sciences. After earning his doctorate in 2022, Yu Bowen joined Alibaba DAMO Academy through Alibaba Group’s highest-level campus recruitment program "Ali Star," serving as an Algorithm Expert (P7). Early in his tenure, he was deeply involved in early training and R&D work for the Tongyi Qwen large model, quickly becoming a core member of the Qwen team, and ultimately the head of post-training. Yu Bowen's departure is closely linked to recent organizational restructuring at Alibaba Tongyi Laboratory. In March, Alibaba Tongyi Laboratory initiated an organizational restructuring, planning to split the originally vertically integrated Qwen team into multiple parallel horizontal modules such as pre-training, post-training, text, and multimodality. This adjustment directly shrank Yu Bowen’s managerial scope and was at odds with his persistent technical approach that "pre-training and post-training must be deeply coupled." Additionally, commercial performance pressure imposed by Alibaba's senior management on the Qwen team further intensified internal disagreements to some extent. On March 3, Yu Bowen submitted his resignation, officially leaving the next day. His position was taken over by former Google DeepMind senior research scientist Zhou Hao. Yu Bowen’s next move also highlights new focal points in large model competition. In recent years, ByteDance’s Seed team has continuously invested in large models and multimodal fields. With Yu Bowen joining as the head of post-training for vision models and multimodal interaction, ByteDance is clearly strengthening its capabilities in the “post-training” phase of multimodality. Post-training, being the key link from general base models to productization and real-world application, directly determines model performance in actual interaction. Yu Bowen’s experience during his time at Qwen—optimizing dialog models, multimodal alignment, knowledge distillation, etc.—fits well with the current technical landscape of the Seed team. Especially in vision and multimodal interaction fields, efficiently fine-tuning and reinforcing learning to make models better “understand” users has become crucial for competitive differentiation among leading companies. Yu Bowen’s move from Alibaba to ByteDance is one example of core talent mobility in this round of AI competition. In January, Qwen Code leader Hui Bin Yuan left Alibaba to join Meta. Earlier, core talent movements had also occurred at international giants such as OpenAI, xAI, and Meta. Behind this wave of talent flow are several deeper changes in the development of the large model industry: First, the era of super technical talents is reshaping the relationship between talent and platforms. The current rapid evolution of large model technology means the judgment and vision of top technical talent have greater influence on technical pathways than ever before. When company strategy deviates from individual technical philosophies, talent tends to seek platforms where they can better fulfill their technical ambitions. Second, computing resources and organizational synergy are now the key reasons for talent retention. Purely financial incentives are no longer sufficient to retain top talent. Whether companies can provide sufficient computing power and build organizational structures aligned with talent’s technical philosophies has become a more crucial factor for retaining core staff. Third, multimodality and post-training are becoming the main battlegrounds for talent in the next stage. As foundational large model abilities begin to converge, how to achieve differentiation through post-training and how to deeply integrate vision and language capabilities have become the focus of leading companies' strategies. Yu Bowen’s decision to join ByteDance’s multimodal team exemplifies this trend. For the industry, the flow of core talent is both a challenge and a catalyst. It forces companies to rethink their approaches to working with top talent, and accelerates cross-platform dissemination and collision of technical ideas. At a time when the large model landscape is still evolving and has not yet settled, the flows of talent are, to some extent, mapping the blueprint of future technological competition. Risk Warning and Disclaimer The market has risks, investment must be cautious. This article does not constitute individual investment advice, nor does it take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any views, opinions, or conclusions in this article are suitable for their particular circumstances. Investing based on this information is at your own risk. ```