Three Waves of AI Talent Migration Reshape China's Large-Model Sector

Baidu, which launched its ChatGPT-style bot Wenxin Yiyan in March 2023, regrouped hundreds of NLP and engineering staff for the effort. That team became the first major target for poaching. Headhunters said at the time that resumes with the keyword Wenxin could draw bidding wars, with offers rising from 2 million to 3 million to 5 million yuan for the same candidate. The poachers were mainly newly founded model startups and existing internet giants, as Alibaba, ByteDance and Tencent all launched their own models between April and September 2023.

A second migration began in early 2025 as competition intensified and DeepSeek disrupted the market. Baichuan switched from general-purpose models to medical verticals, prompting several co-founders and many pretraining engineers to leave. 01.AI moved most of its training and AI infrastructure team into a joint lab with Alibaba Cloud, and top technical leaders departed. ByteDance and Tencent became the main destinations. ByteDance poached across all levels, including former Qwen lead Zhou Chang and former Google DeepMind researcher Wu Yonghui, while Tencent, which had previously been cautious, accelerated hiring for its Hunyuan team and WeChat ecosystem.

According to recruiters in the report, three structural forces drive these flows: share-price pressure, because stock makes up more than half of core researchers' pay; strategic shifts, as seen in Baichuan and Huawei, where model teams are ancillary to hardware sales; and mismatched organizational structures at large, hierarchical companies that slow down AI research. These factors pushed many foundation-model researchers to change jobs or leave large companies altogether.

The third wave is tied to the rise of reasoning and agents. After DeepSeek R1's breakout in early 2025, demand for post-training talent surged, and by 2026 coding and agent teams at ByteDance and Alibaba were described as seeing intense movement. Meanwhile, pretraining recruitment demand fell by 30-40% from its peak, said Zhao Jie, a headhunter. Senior pretraining specialists who left big companies in the past year include Qwen Code leader Hui Binyuan, Tencent Hunyuan pretraining head Yao Xingcheng, Hunyuan multimodal pretraining head Peng Houwen, and ByteDance Seed core expert Gu Quanquan, several of whom moved to startups or more academic-focused research teams.

Few pretraining engineers are converting to post-training, the report said, citing a division of labor and a technical mindset that resists "downward" work. Those who remain at companies such as Moonshot AI's Kimi or DeepSeek often hold valuable option packages. The market increasingly favors researchers in their 20s; some older practitioners have moved to universities or public AI institutes, such as former Kuaishou vice president Wang Zhongyuan, who became president of the Beijing Academy of Artificial Intelligence in early 2024.

The report outlines three current trends. First, the churn rate is accelerating: while model release cycles have compressed from roughly 4.5-6 months to 3.5 months, the average tenure of AI talent in Chinese large-model companies dropped from 2-3 years to 1-1.5 years. Second, AI infrastructure talent is in strong demand, especially engineers skilled in GPU cluster scheduling, FPGA compilation and inference optimization. Third, data roles are moving to center stage. With compute becoming a fixed objective factor and model architecture reaching a platform stage, high-quality data is seen as the industry's core competitive barrier. ByteDance recently restructured several data teams, reflecting this shift.

The report said this talent migration is far from over. As pretraining becomes engineering work and agents and data take center stage, Chinese AI professionals are redefining their career coordinates along the industry's changing competitive rhythm.