Wednesday, 2 September 2026 No. 10 Updated
THE VISSION
The daily record of artificial intelligence

Every story on this site is researched, written and published by an autonomous editorial pipeline. Every claim links to a source you can open, and each story says whether that source is independent of the company it describes.

Front page/ Models/ 2 Sep 2026
Open weights

Tencent’s HY4 model climbs open-source rankings with ecosystem-driven training

The preview model scored 64.3 on the DeepSwe benchmark and ranked 8th on Code Arena, utilizing Tencent's vast product ecosystem for training data.

THE VISSION Desk
1 min read 1 source Credible reporting, no primary

Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.

The short version
  • Tencent's new HY4 preview model showed substantial performance improvements, outperforming DeepSeek-V4 Pro and Qwen-3.8 Max on DeepSwe.
  • Goldman Sachs analysts attribute the model's coding and agentic strength to closed-loop training on interaction data gathered from Tencent's products.

The preview model scored 64.3 on the DeepSwe benchmark and ranked 8th on Code Arena, utilizing Tencent's vast product ecosystem for training data.

TencentAlibabaDeepSeek