Tencent’s Kirin team recently detailed ongoing Hy4 model improvements and their roadmap to train even larger AI models exceeding the current 700 billion parameters. During an extensive 120-minute Reddit AMA, chief AI scientist Yao Shunyu and his team addressed global developer questions about Kirin’s approach and future plans.
Hy4’s Progress and Parameter Scale Debate
When asked why Kirin models remain at around 700 billion parameters while competitors like ByteDance and Kimi push towards 30 trillion to 50 trillion parameters, the team explained that despite smaller model sizes, Kirin’s Hy3 and Hy4 models deliver superior performance with lower costs. The team emphasized that rigorous development of well-optimized smaller models can yield more learnings and better results than simply increasing scale.
However, they acknowledged that Hy4 still has significant room for improvement, particularly through post-training enhancements. The team is heavily investing in large-scale reinforcement learning alongside ongoing efforts to enhance user experience and model safety. Plans to develop even larger models are actively underway.
Future Model Plans and Dense Model Considerations
Addressing questions about the possibility of launching dense models ranging between 24 billion and 31 billion parameters, the Kirin team stated such versions are under consideration. They also indicated a potential release of smaller models aiming to rival the Qwen 27B dense model but are unlikely to produce sub-100 billion parameter models specialized solely for programming tasks.
Yao Shunyu, a Tsinghua graduate and Princeton PhD, joined OpenAI in 2024 before leading Tencent’s large model initiatives. With Tencent’s recent restructuring to strengthen its AI infrastructure and unify multi-modal and language model development under Yao’s leadership, the company aims to push the boundaries of foundational AI model innovation.









