ChainThink reports that on July 6, Tencent officially open-sourced the full version of its 295B-parameter HunYuan 3.0 (Hy3) mixture-of-experts model, changed the open-source license to Apache 2.0, and removed previous geographic usage restrictions on the European Union, the United Kingdom, and South Korea.
The official version also optimizes model capabilities by adding a 3.8B parameter multi-token prediction layer to support parallel generation and reduce inference latency.
Data shows that its hallucination rate decreased from 12.5% to 5.4%, and the error rate in multi-turn interaction tests dropped from 17.4% to 7.9%.
In tool invocation scenarios, Hunyuan 3.0 has been specifically optimized for agent development, reducing cross-framework tool invocation accuracy fluctuations to within 4% on major scaffolds such as Cline and CodeBuddy.
The official release also includes an FP8 quantized version to reduce the VRAM requirements for local deployment and fine-tuning.