Tencent’s Hy4 preview model, trained using data from its extensive ecosystem, has climbed into the top 10 of global AI coding benchmarks, signaling a significant leap in capability and positioning its Hunyuan series as a major player in open-source AI development.
- Hy4 ranks 8th globally on Code Arena WebDev leaderboard.
- Model size increased to 770 billion parameters with 1 million token context window.
- Tencent compresses the model for efficient on-device deployment.
What happened
Tencent Holdings released a Hy4 preview AI model that leverages a closed-loop training approach, drawing on user data from its diverse product ecosystem to enhance model capabilities. This strategy allowed Tencent to deploy the model in real-world applications first, collect valuable interaction data, and iteratively refine subsequent versions.
The Hy4 preview model made a significant impact on global AI coding benchmarks, ranking eighth on the Code Arena WebDev leaderboard and surpassing competitors including Alibaba’s Qwen 3.8-Flash-Next. It also scored higher on the DeepSWE benchmark for software engineering tasks, demonstrating substantial improvements over the previous Hy3 generation.
Why it matters
Tencent’s integration of ecosystem-driven training represents a strategic advancement in AI development, enabling the company to tailor AI models more effectively for productivity and coding-related workloads. This approach positions Tencent’s Hunyuan AI series as a leading contender in the open-source AI space, strengthening its competitive edge amid China’s rapidly evolving AI landscape.
The model’s scale—770 billion parameters and an expanded 1 million token context window—along with its optimized compression to only 214 gigabytes, offers an appealing mix of performance and deployment efficiency. This balance is critical as enterprises increasingly require advanced AI capable of running on various hardware platforms with manageable infrastructure costs.
What to watch next
Tencent plans to launch a full commercial version of the Hy4 series later this year, following the preview phase. Market watchers will be observing how effectively the model can address current limitations such as longer reasoning times and excessive self-verification, which are expected to improve through continued training and real user feedback.
The company’s substantial capital expenditure increase in AI infrastructure—almost tripling year over year to 52.8 billion yuan in Q2—suggests strong ongoing commitment. This investment, coupled with Hy4’s advancements, could further elevate Tencent’s position as a dominant AI provider within China and the broader global market.