Tencent Hunyuan 3.0 Official Open Source Version: Switching to Apache 2.0 removes overseas restrictions, halved the hallucination rate

Golden Finance reports that on July 6, Tencent officially released the 295B parameter version of Hunyuan 3.0 (Hy3) Mixture of Experts (MoE) model. The most significant change is that the open-source license has been switched to the most permissive Apache 2.0, removing the previous geographic restrictions prohibiting use in the European Union, the United Kingdom, and South Korea.

When the preview version was first released in April this year, many overseas or multinational teams were discouraged by strict geographic restrictions and a 100 million monthly active user commercial red line. The official version has cleared all compliance hurdles and made hardcore optimizations for deployment challenges: it integrates a fast-slow thinking mechanism and incorporates a 3.8B parameter Multi-Token Prediction (MTP) layer for parallel generation, effectively reducing inference latency. Through fine-grained data cleaning and training constraints, the official version has reduced the hallucination rate from 12.5% to 5.4%, and the error rate in multi-turn interaction tests has been cut from 17.4% to 7.9%.

To address the persistent issue of tool calls going off-track during agent development, Hunyuan 3.0 has also made targeted improvements. Whether in mainstream frameworks like Cline or CodeBuddy, the fluctuation in cross-framework tool call accuracy is kept within 4%, ensuring more stable output. The simultaneously released FP8 quantization version also lowers the GPU memory threshold for local deployment and fine-tuning.