Tencent Hunyuan has hired researcher Tian Yunlong from OpenAI to head its multimodal models team. He will report to Chief AI Scientist Yao Shunyu. This move is part of a recent major campaign to attract talent in the field of artificial intelligence.
Tian's educational background follows the typical trajectory of elite AI specialists in China. He earned a bachelor's degree from Tsinghua University, followed by a master's degree from the Chinese University of Hong Kong under the guidance of renowned computer vision pioneers Tao Xiaowu and Wang Xiaogan. He completed his studies in EECS at the Massachusetts Institute of Technology (MIT) under the direction of Philip Isola, inventor of CycleGAN.
According to Tian's Google Scholar profile, he has 28,778 citations and an h-index of 35, with 3,436 citations recorded just for 2026. After obtaining his Ph.D., Tian worked in the research team at Cambridge Google DeepMind before moving to OpenAI as a research scientist.
At OpenAI, he specialized in machine perception, generative models, and visual information representation. His prior collaboration with Yao Shunyu at OpenAI makes their subsequent partnership at Tencent a natural progression. One of Tian's most significant academic contributions is the Supervised Contrastive Learning method, which extended the framework of contrastive learning for supervised scenarios by clustering samples of the same class in the representation space. This work influenced subsequent designs of vision pre-training models, including the SimCLR and MoCo series.
More recent work by the team, Fluid (October 2024), demonstrated that autoregressive text-to-image generation models using continuous tokens significantly outperform discrete token approaches, with random order generation surpassing fixed raster order generation. The team trained the Fluid 10.5B model as a new benchmark.
Tian's addition is part of a series of changes within the Hunyuan multimodal technology team. Following the departure of former Head of Multimodal Direction Liu Wei at the end of 2024, Tencent established independent LLM and multimodal model departments within TEG in April 2025. This was followed by recruitment from Microsoft Research Asia and Alibaba's Tongyi lab. Nevertheless, Hunyuan's capabilities in multimodality are often considered lagging behind ByteDance's Doubao and Alibaba's Tongyi Qianwen.
Tian's deep expertise in generative models and visual representation, combined with Yao's specialization in post-training language models, provides comprehensive coverage of core LLM and multimodality areas. For Tencent, Tian's arrival signals an intent to create a genuine technological differentiator in multimodality, rather than merely remaining a tool for integration into the WeChat and QQ ecosystem. The company expects that high-caliber research personnel will allow it to close the gap with competitors that have larger funding.
Gu Yushen, a Tsinghua graduate student who holds a special scholarship and has nearly 5000 citations, has joined DeepSeek in anticipation of the official release of DeepSeek V4, scheduled for mid-July.
DeepSeek is actively hiring specialists in algorithmic research, engineering, product development, operations, and data engineering in preparation for the launch of DeepSeek V4. Among the authors mentioned in the DeepSeek V4 scientific paper is Gu Yushen. He is a candidate of sciences from Tsinghua University and received the Tsinghua Graduate Special Scholarship for 2025, according to a report by Machine Intelligence.
Gu Yushen has officially joined DeepSeek, adding one of the most promising young AI researchers from Tsinghua to the company's staff. Previously, he received the Apple PhD Scholarship for 2025 and the Ant Group In-Tech Scholarship. He is a graduate of the PhD program in Computer Science at Tsinghua University, where he also completed his undergraduate studies. His research is conducted under the supervision of Professor Huang Minli in the Conversational AI group.
Gu's research focuses on improving efficiency throughout the entire lifecycle of large language models. His work covers pre-training, subsequent adaptation, and inference. He concentrates on three main areas: data selection for pre-training, where he develops theories and algorithms to optimize data selection for creating more capable and efficient models; model compression through knowledge distillation, where he designs methods for transferring knowledge from large models to smaller, more deployable ones; and developing efficient model architectures by exploring new structures that reduce computational costs while improving performance.
According to his Google Scholar profile, his total citation count reaches nearly 5000. Two of his works have exceeded the threshold of 1000 citations each: 'Pre-trained Models: Past, Present and Future' and 'MiniLLM: Knowledge Distillation of Large Language Models'. He has been the first author on several papers accepted at leading AI conferences such as NeurIPS, ICLR, ACL, and others.
Notably, Gu made a significant contribution to the series of hybrid architectural language models, Jet-Nemotron. This series demonstrates state-of-the-art accuracy with full attention while maintaining excellent efficiency. The 2-billion parameter version of Jet-Nemotron outperformed Qwen3, Qwen2.5, Gemma3, and Llama3.2 on MMLU and MMLU-Pro benchmarks, while achieving up to 53.6 times faster generation speed on an H100 GPU with a context length of 256K, surpassing larger full-attention MoE models, including DeepSeek-V3-Small and Moonlight.
As Gu himself noted, 'When hardware resources are limited, algorithmic innovations become the key to overcoming computational bottlenecks.' His research philosophy aligns with DeepSeek's emphasis on developing efficiency-oriented AI, making him a strategic acquisition for the team ahead of the V4 launch.
