Hy3 preview

Hy3 preview - Open-sourced agentic MoE

Hy3 preview is Tencent’s open-sourced flagship model in the Hy series, positioned for real-world agent workflows with long-context understanding and fast-and-slow thinking.

Defining Characteristics

• Mixture-of-Experts architecture: Tencent describes Hy3 preview as a MoE model with 295B total parameters and 21B activated parameters per token.
• 256K token context window: Official Tencent documentation states support for up to a 256K token context window.
• Three inference modes: Public listings mention configurable reasoning levels such as disabled/no_think, low, and high to balance speed and depth.
• Agent-oriented workflow support: Tencent says the model is designed for agent capabilities, complex workflows, and MCP toolchain orchestration.

Core Design

• Tencent highlights a fast-and-slow thinking fusion design and substantial improvements in inference efficiency compared with earlier infrastructure.
• The model is open-sourced and made available on GitHub, Hugging Face, ModelScope, and GitCode, with support for vLLM and SGLang deployment.

Deployment

• Hy3 preview is integrated into Tencent products such as Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, and Tencent LearnShare.
• Tencent Cloud TokenHub and OpenRouter both list the model, with limited-period free access reported on limited-period free access reported on OpenRouter.