Tencent Hunyuan 3.0 Official Version Open-Sourced: Switched to Apache 2.0 to Clear Overseas Restrictions, Halving the Hallucination Rate

According to Beating monitoring, Tencent has officially released the stable version of Hunyuan 3.0 (Hy3), a 295B-parameter Mixture of Experts (MoE) model. The most core change is that the open-source license has been switched to the most permissive Apache 2.0, abolishing the previous regional terms that prohibited use in the European Union, the United Kingdom, and South Korea.

When the April preview version was released this year, many overseas or multinational teams had to hesitate because of strict regional restrictions and the red line of 100 million monthly active users for commercial use. The stable version removes all compliance obstacles and also delivers hardcore tuning for real-world deployment challenges: it combines fast-and-slow thinking mechanisms and adds a 3.8B-parameter Multi-Token Prediction (MTP) layer for parallel generation, effectively lowering inference latency. Through fine-grained data cleaning and training constraints, the stable version has reduced the hallucination rate from 12.5% to 5.4%, and the error rate in multi-turn interaction testing has also been cut from 17.4% to 7.9%.

To address the persistent issue that tool calls during agent development can easily go off track, Hunyuan 3.0 also provides targeted enhancements. Whether in popular scaffolding frameworks like Cline or CodeBuddy, fluctuations in tool-call accuracy across frameworks are kept within 4%, resulting in more stable outputs. The FP8 quantized version released alongside by the official team also lowers the GPU memory threshold for local deployment and fine-tuning.