Dynamic Vision Beating AI Newsletter: Tencent has released the Hy4 preview with a new model utilizing the MoE architecture. The total parameters are 770 billion, with 49 billion parameters activated per inference, and the context window expanded to 1 million. Compared to Hy3 with 295 billion total parameters, 21 billion activation parameters, and a 256K context, this generation has doubled in scale directly, with the long context expanding to around 4 times.

This time, Tencent has not only emphasized general benchmarking but has specifically strengthened real-world tasks such as coding, office work, gaming, and research. For example, the model can build a Three.js 3D website from scratch, create playable game demos in Unity, and process 72 financial documents at once to check for duplicate reimbursements, over-limit amounts, and budget anomalies. Hy4 is also trained and iterated with Tencent products like WorkBuddy and CodeBuddy.

A particularly unique aspect is that Hy4 has begun to engage in its own R&D. Tencent stated that it will help optimize training methods, data strategies, evaluation systems, and underlying operators, propose solutions, conduct experiments, and then adjust based on the results. The experiment-generated code, logs, and feedback will enter the next round of R&D, forming an initial self-enhancement cycle.

The Hy4 preview has been integrated with WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter. Tencent had previously hinted in its financial report that Hy4 would be larger than Hy3 and that real product feedback would be a key source for model training.