According to CoinMeta, Tencent's Hunyuan Hy4 model adopts a MOE architecture with a total of 770 billion parameters, activating 4.9 billion of them each time, and its context expansion reaches up to 1 million. Compared to Hy3, the total parameters of Hy4 have doubled, and its long-context capacity has increased by about 4 times. Tencent emphasizes the model's application capabilities in real tasks such as coding, office work, gaming, and scientific research, such as building a three.js 3D website from scratch, creating demo games within Unity, as well as processing 72 financial documents, checking for duplicate reimbursements, and budget anomalies. Hy4 also trains and iterates together with Tencent's other products like workbuddy and codebuddy, forming an initial self-improvement cycle. Hy4 and preview have been integrated into multiple Tencent products. Previously, Tencent announced in its financial reports that Hy4 would be larger than Hy3, and it uses feedback from real products as an important source for model training.