Tencent Holdings has launched Hy4, its most powerful AI model to date, marking the first major release under former OpenAI researcher Yao Shunyu, who joined the Chinese tech giant to lead its foundational AI development efforts.
The open-source model, announced on Thursday, uses a mixture-of-experts architecture with 770 billion total parameters — but only 295 billion active at any given time. That design choice deliberately bucks the industry's recent obsession with trillion-parameter models, instead prioritizing efficiency and real-world business applicability.
The model is on par with top Chinese models from competitors but still trails flagship products from OpenAI and Google DeepMind, according to the company. Tencent's previous flagship, HY 2.0, released in early December, had over 400 billion parameters — making Hy4 paradoxically smaller yet more capable.
The model was developed through close collaboration between Tencent's foundational model team, Hunyuan, and its Yuanbao AI application team, the company said. 'By seamlessly aligning product-side requirements with underlying technology, we have successfully bridged the gap between model capability and user value.'
Agentic capabilities — the ability of AI systems to autonomously complete multi-step tasks — were described as 'one of the most significantly improved areas' in the release, with strong performance on in-house benchmarks testing compatibility with the agentic AI tool OpenClaw.
Hy4 is already deployed across Tencent's flagship AI products, including its consumer app Yuanbao and the CodeBuddy coding assistant. The model's smaller active footprint could make it significantly cheaper to serve at scale, a key advantage as Chinese AI companies race to make their models commercially viable.
The release comes amid an intensifying AI model race in China, where Tencent competes with DeepSeek, Moonshot AI, Baidu, and Alibaba for dominance in the world's second-largest AI market.
