Tencent has officially released Hy3, an open-source artificial intelligence model built on a Mixture-of-Experts architecture containing 295 billion total parameters. While 21 billion parameters remain active at any given time, the system includes an additional 3.8 billion for a specific MTP layer. The model supports context lengths of up to 256,000 tokens. Tencent states that Hy3 matches the performance of models two to five times its active size. A blind evaluation by 270 experts saw the model score 2.67 out of 4, beating GLM-5.1 which scored 2.51. Internal testing showed the hallucination rate dropped from 12.5 percent to 5.4 percent.
The release is significant because it offers a high-capacity model under an Apache 2.0 license, allowing broad adoption without restrictive terms. The FP8-quantized version improves efficiency for deployment on standard hardware. Plans include support for OpenRouter and Cline, expanding accessibility beyond Tencent’s own ecosystem. The company has already integrated the model into WorkBuddy, Yuanbao, WeChat, and the game assistant for Path of Exile: Advent.
- Apache 2.0 licensing enables unrestricted commercial use
- Context window reaches 256,000 tokens
- Integrated into WeChat and Path of Exile: Advent
