Tencent Hunyuan 3.0 Reaches Formal Release with Apache 2.0 License, Cuts Hallucination Rate to 5.4%

According to Beating monitoring, Tencent formally released Hunyuan 3.0, a 295-billion parameter mixture-of-experts (MoE) model, with the most permissive Apache 2.0 open-source license, removing previous geographic restrictions that prohibited use in the EU, UK, and South Korea.

The formal release eliminates compliance barriers that hindered overseas and cross-border teams since the April preview version. The model reduces hallucination rates from 12.5% to 5.4% through fine-grained data cleaning and training constraints, while multi-turn interaction error rates dropped from 17.4% to 7.9%. Tool calling accuracy remains stable across frameworks with fluctuations within 4 percent.