Hy3

The official Tencent Hy team model card on Hugging Face documents Hy3 preview as a 295B-parameter Mixture-of-Experts model with 21B activated parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team and described as the first model trained on the team's rebuilt infrastructure. Architecture details inc The model card highlights strong results across three areas: STEM and reasoning benchmarks such as FrontierScience-Olympiad, IMOAnswerBench, the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26), and the China High School Biology Olympiad (CHSBO 2025); context learning and instruction following measured on

A Tencent Cloud community article frames Hy3 preview as the first step of the Hunyuan rebuild, reiterating that the model was released and open-sourced on April 23, 2026. It repeats the headline specs: a hybrid fast-and-slow-thinking Mixture-of-Experts design with 295 billion total parameters, 21 billion active paramet The post restates Tencent's three principles for the rebuild: systematic capabilities that avoid lopsided models covering reasoning, long-form understanding, instruction following, dialogue, code, and tools; authentic evaluation that steps away from easily gamed leaderboards; and tight model-and-inference co-design wit

A Baidu Baike encyclopedia entry describes Hy3 preview as an artificial intelligence model released and open-sourced by Tencent's Hunyuan team on April 23, 2026, following a February reorganization. The model uses a Mixture-of-Experts architecture with 295 billion total parameters, 21 billion activated parameters, and The article reports that Token usage for Hy3 preview reached roughly ten times that of the previous-generation Hy2 model, with significant usage in code and Agent-type scenarios, and cites improvements in PowerPoint generation tasks of 20 percent shorter duration, 20 percent higher task success rate, and 10 percent hig