by TencentCN
Build an open-weight foundation model efficient enough for wide adoption while remaining competitive with closed frontier models, and drive Tencent Cloud adoption through developer ecosystem growth.
Released July 8, 2026 under Apache 2.0. Hy4 multimodal in training.
Updated 8 Jul 2026Checked 11 Oct0 updates this week
Hy3 uses a sparse mixture-of-experts architecture, activating only 21B of its 295B parameters per token, reducing compute while maintaining performance.
Hy3 includes a 3.8B-parameter multi-token prediction layer for improved inference efficiency and speed.
Trained on a diverse corpus covering many languages and specialized domains, with emphasis on coding and reasoning tasks.
Hy3 is released as open weights under Apache 2.0, allowing free use and commercial deployment without restrictions.
| Spec | Hunyuan Hy3 |
|---|---|
| Total parameters | 295 billionR (reported) |
| Active parameters per token | 21 billionR (reported) |
| Context window | 256K tokensR (reported) |
| Claimed capability vs parameter scale | Matches models 2–5x largerR (reported) |
R reported by the company
| Spec | Hunyuan Hy3 | Reflection AI Beam |
|---|---|---|
| Total parameters | 295 billionR (reported) | 501 billionR (reported) |
| Active parameters per token | 21 billionR (reported) | 23 billionR (reported) |
R reported by the company
Tencent
Hunyuan, messaging, cloud, WeChat, QQ, gaming, social
Tencent is a Chinese technology conglomerate that operates WeChat (1.3B+ users), QQ, major games (League of Legends, Honor of Kings), and Tencent Cloud. It released Hunyuan Hy3, a 295B open-weight mixture-of-experts foundation model under Apache 2.0 in July 2026.
