Tencent Launches Hy4: A 770B Parameter Open-Weight MoE Model with Native Computer-Use and 1M Context

Tencent has officially announced the full production release of Hy4, establishing a formidable new benchmark in the open-weights and frontier AI ecosystem. Built on a large-scale Mixture-of-Experts (MoE) architecture, Hy4 encompasses 770 billion total parameters while activating an efficient 49 billion parameters per token during inference. This sparse activation strategy is designed to deliver frontier-grade reasoning and generation depth while maintaining manageable throughput speeds for enterprise hosting. A cornerstone of the Hy4 architecture is its native support for a 1-million-token context window, accompanied by an extensive 65,536-token generation limit. This allows the system to process massive multi-repository codebases, extensive financial filings, and long-horizon multi-step instructions without context fragmentation. Moving beyond earlier text-focused iterations, the full release incorporates native multimodal vision, showing competitive accuracy in technical diagram parsing, visual mathematics, and complex data visualization analysis. Diverging from purely conversational models, Tencent has explicitly oriented Hy4 toward autonomous computer-use and agentic capabilities. The model features specialized training to operate across command-line interfaces and graphical desktop operating systems, executing multi-turn tool calling across professional creative and development environments. Demonstrations highlight its ability to generate and manipulate assets across CAD suites, Blender, and modern 3D game engines including Unreal Engine, Unity, and Godot. The model is rolling out across developer channels with planned weight releases and system cards, alongside hosted API endpoints via Tencent Cloud and third-party routing platforms like OpenRouter. As the self-hosted community continues to seek high-parameter open architectures compatible with modern inference runtimes like vLLM, Hy4 represents one of the most capable open-weight agent foundations to date. ~AI ~OpenSource ~MachineLearning ~Tencent