Tencent's Hy4 climbs open-source AI leaderboard thanks to closed-loop training
Tencent's Hy4 preview model ranked eighth globally on Code Arena's WebDev leaderboard, surpassing Alibaba's Qwen and DeepSeek. The company's strategy of deploying preview models across its product suite to collect user data and feed back into training gave it an edge. Goldman Sachs analysts view this closed-loop approach as particularly relevant for productivity and coding workloads.
The Hy4 preview's jump from 34th to 8th place on Code Arena's WebDev leaderboard marks a significant leap for Tencent's Hunyuan series in competitive coding benchmarks. On the DeepSWE evaluation, Hy4 scored 64.3, edging out DeepSeek-V4 Pro's 62.7 and Alibaba's Qwen-3.8 Max at 56.6. The model was released Friday and immediately demonstrated improved coding capabilities over its predecessor, Hy3.
Goldman Sachs analysts, led by Ronald Keung, attribute this performance to Tencent's "product-plus-model" strategy. By first deploying preview versions across its product ecosystem, the company gathers real-world user interaction data that feeds back into subsequent training cycles. This closed-loop system appears particularly effective for productivity and coding workloads, where authentic task trajectories and evaluation signals help differentiate models in the agentic AI era.
This closed-loop training approach could reshape how AI models are developed, as real-world usage data becomes as valuable as curated training sets. Smaller developers without vast product ecosystems may struggle to compete, potentially consolidating AI leadership among platform giants. For businesses and individual developers relying on open-source models, Tencent's rise could mean more capable coding assistants, but also greater dependence on companies that control both the products and the training pipelines feeding them.