Tencent has quietly launched its most capable AI model to date, Hy3 preview. This new model boasts a 295 billion parameter Mixture-of-Experts architecture, with only 21 billion active parameters per query, significantly reducing operational costs compared to competitors.
On the SWE-bench Verified coding benchmark, Hy3 preview demonstrated a substantial 40% improvement, jumping from 53% with its predecessor, Hy2, to 74.4%. This performance positions it competitively against leading models like Claude Opus and GLM-5.
Hy3 preview also shows marked advancements in agentic capabilities and web browsing tasks. On BrowseComp, its score rose from 28.7% to 67.1%, indicating enhanced ability in complex web research without human guidance.
In reasoning, the model achieved top scores among Chinese competitors on a real-world math PhD qualifying exam, scoring 88.4. The model was developed rapidly, with training and release completed in under three months, attributed to an infrastructure overhaul led by chief AI scientist Yao Shunyu.
The Hy3 model is already integrated into Tencent's ecosystem, including Yuanbao and QQ. API access is available on Tencent Cloud starting at approximately $0.18 per million input tokens.