🧠 Food for thought

Implications, context, and why it matters.

Tencent’s Hy3 marks a broader reset

• Former OpenAI researcher Yao Shunyu joined Tencent to lead Hy3, bringing a more practical approach to model design 1.
• Yao’s team rebuilt training systems to improve stability and tackle business needs across Tencent’s products, rather than chasing benchmark scores alone 1.
• Hy3 uses a Mixture of Experts (MoE) architecture, which turns on only part of the model for each task. It has 295 billion total parameters and uses 21 billion at a time to curb running costs 1.
• In tests, Hy3 carried out complex research on its own, turned data into visualizations, and built a working web game from a simple prompt 1.

China’s AI companies are building models while weighing outside bets

• Tencent released Hy3 while discussing a place in DeepSeek’s first outside funding round, including a proposal for up to a 20% stake. The talks are still underway and remain unsettled 2.
• Tencent and Alibaba are each building their own models while looking at rivals 2. That puts them across China’s fast-moving AI market through funding, plus cloud infrastructure 2.
• Each company also runs a large cloud business and aims to supply computing power to AI startups that rely on data centers to scale their platforms 2.
• That route gives big tech groups a hedge, with a chance to benefit from advances that come from internal teams or backed startups if deals happen 2.