Arena (LMArena)

Arena (LMArena) has published an Alignment Index leaderboard based on more than 90,000 real agent sessions across 27 models, measuring overreach, misattribution, and false-completion failure signals. Tencent Hunyuan Hy4 Preview scored 78.5, ranking fifth globally and first in China. Its 1.47% overreach rate was the lowest of any Chinese lab, while its 13.24% false-completion rate remained above the 10% average.

The model was led by Yao Shunyu, who took charge of Tencent's AI Data department in September, and was jointly designed with WorkBuddy and CodeBuddy. Tencent Hunyuan Hy4 Preview is available through OpenRouter and Tencent Cloud TokenHub.