Hy3: Tencent's New 295B Open-Weights Mixture-of-Experts

Tencent releases Hy3, a 295B open-weights Mixture-of-Experts AI model under Apache 2.0 license, claiming it outperforms GLM-5.1 in expert blind tests.

Successor to a preview released in April, Hy3 expands Tencent's Hunyuan family: a 295-billion-parameter Mixture-of-Experts architecture, distributed as open weights under an Apache 2.0 license and open for commercial use. The publisher positions it as the best in its size class and claims it rivals reference models two to five times larger.

The claimed advancements primarily concern agentic and productive uses: code, office automation, financial modeling, frontend design, and game development. In a blind evaluation conducted by 270 experts on their own tasks, Tencent places Hy3 at 2.67 out of 4, ahead of GLM-5.1 (2.51), with the clearest gap in frontend, CI/CD, and data management. The company also highlights, based on its internal measurements, an hallucination rate reduced from 12.5% to 5.4% and a multi-turn conversation incident rate decreased from 17.4% to 7.9%.

Regarding access, the model is available on GitHub, Hugging Face, ModelScope, and AtomGit, with an API priced at 1 yuan per million input tokens and 4 yuan per million output tokens, and a two-week free trial via OpenRouter.