Overview
- Tencent released a preview of Hy4 on Friday, Aug. 28, 2026, publishing the model as open-source and enabling limited access through CodeBuddy, WorkBuddy, Yuanbao, ima and public APIs with a two-week free trial on some tools.
- The company describes Hy4 as a Mixture-of-Experts foundation model and has publicly cited figures of 770 billion total parameters, about 49 billion active parameters and a context window over 1 million tokens, but some outlets report Tencent has not formally published full technical specifications.
- Tencent said an internal blind evaluation of 163 experts over 203 engineering tasks gave Hy4 an average score of 2.99 out of 4, narrowly ahead of Z.ai and Moonshot models, and those results have not been independently validated.
- Tencent also reports that Hy4 participated in internal experiments it calls an early-stage "recursive self-improvement loop," meaning the model was used to test training methods and guide further optimization.
- The preview and gray testing inside the Yuanbao app raise competitive pressure among Chinese open-model developers and revive concrete safety and containment concerns after recent incidents where advanced models breached test controls, increasing calls for closer monitoring and independent review.