Tencent appoints ex-OpenAI researcher as chief AI scientist

Tencent has reorganized its AI operations and promoted Vinces Yao Shunyu, a former OpenAI researcher, to chief AI scientist under the CEO’s office.

Yao now reports directly to Tencent president Martin Lau and is responsible for leading new AI infrastructure projects, including large language model development.

As part of the restructuring, Tencent’s Technology Engineering Group created two new departments: AI Infra, focused on large-model training and inference platforms, and Data Computing Platform, which will develop big data and machine learning services.

Yao will head both the LLM development and AI Infra teams, reporting to TEG president Lu Shan.

The company also renamed its data platform team to AI Data and dissolved the Machine Learning Platform department, with staff moving to the new divisions.

At Tencent’s recent earnings call, president Martin Lau said the company is hiring more AI researchers, and Tencent has open-sourced over 30 new AI models this year.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Tencent’s AI reshuffle leaves open questions on compute and scale

Q3 capital expenditure (CapEx) was RMB 13 billion, down 24% year over year, with 2025 CapEx expected below prior guidance 1. That hints at tight AI infrastructure spend versus typical hyperscaler budgets (hyperscalers are the largest cloud providers).
Management said external cloud growth is capped by “AI chip supply constraints” and current GPU stock is “sufficient for internal use” 1. No numbers on GPUs, cluster size, or compute capacity to judge if Tencent can train competitive foundation models (large and general-purpose AI systems) at scale.
Tencent calls its Hunyuan imagery and 3D generation models “industry-leading globally” 2. Management also said “no decisive better AI model in China” 1.

Open-source teams can use Tencent’s 30+ models for niche apps

Available releases include HunyuanVideo with 13 billion parameters for video generation 3, Hunyuan-Large with 389 billion parameters as the largest open-source Transformer-based MoE (Mixture of Experts) model 4, plus Hunyuan3D 2.0 for high-resolution 3D assets 5. They support text-to-video, text-to-3D, and 256K-token context.
SaaS builders can plug in HunyuanWorld 1.0 for 3D worlds from text or images 6. Hunyuan-A13B uses an 80 billion parameter MoE for resource-limited use 7. It supports FP8 or INT4 quantized inference, and runs on TensorRT-LLM or vLLM 7.
Before shipping, check each repo’s LICENSE for rights and limits. Hunyuan-1.8B includes licensing details 8. Teams should confirm fit by use case, since management said different models suit different scenarios 1.