On the DeepSWE benchmark, the Hy4 preview scores 64.3, surpassing Alibaba’s Qwen-3.8 Max at 56.6 and DeepSeek-V4 Pro at 62.7
Tencent Holdings’ strategy of using its vast product ecosystem to train its new Hy4 preview model gives it an edge in developing AI agents and brings its flagship model suite back into the top tier of open-source offerings, according to analysts.
The Chinese tech giant’s “differentiated product-plus-model strategy”, where preview models were first deployed across Tencent’s suite of products, enabled it to collect user data before feeding the information back into subsequent rounds of training, Goldman Sachs analysts said in a research note on Monday.
“We view this closed-loop approach as particularly relevant for productivity and coding workloads, where real-world task trajectories, user interactions and evaluation signals drive model differentiation in the agentic AI era,” said the analysts led by Goldman’s head of Asia internet research Ronald Keung.
They added that the Hy4 preview brought Tencent’s Hunyuan series back to the forefront of open-source models, with notable gains in coding capabilities over Hy3.
Released on Friday, Hy4 preview ranked eighth globally on Tuesday on Code Arena’s WebDev leaderboard – a real-time competition evaluating large language models on real-world coding tasks – just behind Anthropic’s Claude Fable 5 and outperforming Alibaba Group Holding’s Qwen 3.8-Flash-Next in ninth place.
By comparison, Tencent’s previous generation, Hy3, ranked 34th on the same benchmark on Tuesday.