Tencent Hunyuan's Multimodal AI Chief Hu Han Departs as Team Pivots to World Models

Hu Han, head of multimodal understanding at Tencent Hunyuan, has resigned to found his own startup, the Chinese publication Intelligent Emergence (36Kr, 智能涌现) reported on July 23, 2026. According to the publication, Tencent plans to redirect his former research group to the study of world models.

Who is Hu Han

Before Tencent, Hu Han was a principal researcher in the visual computing group at Microsoft Research Asia. He joined Tencent in early 2025 and led research on large vision models. After an internal reorganization, he moved to the "Frontier" advanced research group within the large language model department and began reporting to Yao Shunyu.

According to 36Kr, Hu Han was also responsible for developing world models — systems that simulate physical reality rather than simply describing an image. The publication believes this is the direction that may become the new focus of his former team.

Why Tencent is rethinking multimodal

Tencent is reallocating resources toward foundation language models, and multimodal understanding research is losing priority. Accuracy in recognizing text, images and video has already exceeded 85% in this field, so further investment yields limited returns, a researcher cited by 36Kr points out.

Yao Shunyu, a former OpenAI researcher who now heads Tencent's large language model department, plays the key role in the shift. After he arrived, the company disbanded AI Lab and folded its core developers into the LLM department. It also stepped up hiring, notably from ByteDance's Seed Infra and post-training teams.

• Hu Han has resigned and is founding a startup (36Kr report, July 23, 2026)
• Previously a principal researcher at Microsoft Research Asia; at Tencent since early 2025
• His replacement on the VLM track is Tian Yunlong, a former OpenAI researcher with a PhD from MIT, who joined Tencent in early July 2026
• Tencent's capital expenditure for 2025 was 79.2 billion yuan
• For comparison: Alibaba spent 126 billion yuan in the fiscal year ending March 2026; ByteDance plans up to $70 billion on AI infrastructure in 2026 (Bloomberg data)

Recognition or monetization: which matters more?

Multimodal understanding is hard to monetize directly, and that is one reason for the change in priorities. Users are not willing to pay for scenarios like image recognition or "describe this image." They do pay for document processing and for creating presentations and reports, because those rely on the model's reasoning, coding and agentic capabilities.

"No user is willing to pay for image recognition — the market is full of free alternatives," a product manager for the

Yuanbao app told 36Kr, explaining the logic behind the review.

Yao Shunyu enforces strict discipline in his new team. A month before the release of the Hy3 model, when problems surfaced with part of the data, he warned the team, according to the publication LatePost:

"Data is critically important.

If this happens again, you're fired immediately," — Yao Shunyu, head of Tencent's large language model department.

What this means

Tencent is deliberately slowing down multimodal understanding, a mature but weakly monetizable field, and pulling talent and compute toward foundation models and frontier topics like world models. The first result of the strategy is the Hy3 model, released on July 6, 2026, which 36Kr says already competes with GLM-5.2 and DeepSeek V4 Pro in the mid-size class.

Frequently asked questions

What are world models and why is Tencent betting on them?

World models are models that learn to simulate the physical world and its dynamics, rather than just recognizing what is in an image. Tencent considers this direction more promising than mature multimodal understanding, and according to 36Kr, Hu Han's former group may switch to it.

Who will replace Hu Han at Tencent?

The vision-language model (VLM) track will be led by Tian Yunlong, a former OpenAI researcher with a PhD from MIT, who joined Tencent's large language model department in early July 2026 and reports to Yao Shunyu.

What is the Hy3 model?

Hy3 is a Tencent large language model released on July 6, 2026, and Yao Shunyu's first major result at the company. According to 36Kr, it is comparable to GLM-5.2 and DeepSeek V4 Pro in the mid-size class.