Tencent Releases 770-Billion Parameter Open-Source AI Model Hy4 Preview

Tencent released a preview version of a new open-source AI model aimed at tasks such as software engineering, research and financial analysis on Hugging Face. [1] The model, called Hy4 preview, utilizes a "mixture-of-experts" design with 770 billion total parameters. [1] However, only about 49 billion parameters are activated for any given text request. [1][2] This architecture spreads capacity across 78 layers, utilizing 256 routed experts and one shared expert. [2]

The model features a context window exceeding one million tokens, designed for multi-step work. [2] Tencent's Hunyuan team stated they invited colleagues from software engineering, game development, finance, and security to construct real-world task data for the model. [2] The Chinese technology giant indicated it plans to integrate the model alongside Tencent products. [1] Tencent claims Hy4 delivers top-tier performance among open-source models on real-world productivity tasks, though independent verification is still pending. [2]

Tencent's release of Hy4 signals a strategic move to compete in the open-source AI space by focusing on practical, enterprise-oriented applications rather than raw benchmark supremacy. By using a Mixture-of-Experts architecture, Tencent achieves massive model capacity while keeping computational costs manageable during inference, a technique increasingly popular among leading AI developers. The focus on real-world data from internal business lines suggests an intent to create highly specialized, production-ready tools. What the sources don't address: How the open-source community will adopt and fine-tune Hy4 compared to other leading open models like Meta's Llama series or Mistral's offerings.