Tencent's Hy4 model climbs open-source AI rankings via ecosystem training

Tencent Holdings has made significant strides with its Hy4 preview model, achieving a score of 64.3 on the DeepSWE benchmark, which outperforms Alibaba's Qwen-3.8 Max (56.6) and DeepSeek-V4 Pro (62.7). Released on Friday, Hy4 secured the eighth position globally on Code Arena's WebDev leaderboard, trailing Anthropic's Claude Fable 5 and preceding Alibaba's Qwen 3.8-Flash-Next at ninth. This represents a remarkable leap from Tencent's earlier model, Hy3, which was ranked 34th. Analysts, including Goldman Sachs' Ronald Keung, noted that Hy4's success stems from Tencent's unique product-plus-model strategy, leveraging user data from its extensive ecosystem for continuous training, particularly enhancing productivity and coding tasks in the evolving agentic AI landscape.