Tencent just open-sourced Hy4 preview, a new mixture-of-experts model with 770 billion total parameters, 49 billion active parameters, and a 1 million-token context window. That's more than double the scale of Hy3 (295 billion total, 21 billion active, 256K context) and roughly four times the long-context capacity.
Tencent isn't leading with generic benchmark scores this time. Instead, it's emphasizing real-world tasks: coding, office work, gaming, and research. The model can build a Three.js 3D website from scratch, put together a playable game demo in Unity, and process 72 financial documents at once to catch duplicate reimbursements, spending limits that were blown past, and budget anomalies. Hy4 was trained and iterated together with Tencent's own products, including WorkBuddy and CodeBuddy.
The more unusual part: Hy4 is already taking part in its own development. Tencent says the model helps optimize training methods, data strategies, evaluation systems, and low-level operators — proposing experiments, running them, then adjusting based on the results. The code, logs, and feedback from those experiments feed back into the next round of R&D, forming an early self-improvement loop.
Hy4 preview is already plugged into WorkBuddy, CodeBuddy, Yuanbao (Tencent's AI assistant), ima, Tencent Cloud TokenHub, and OpenRouter. Tencent had already teased in its earnings report that Hy4 would be bigger than Hy3, and it looks like real product feedback is doing a lot of the heavy lifting.