Tencent Hunyuan released and open-sourced Hy4 preview, a Mixture-of-Experts model with 770 billion total parameters and 49 billion activated parameters per token. The model supports context lengths exceeding 1 million tokens.
Tencent claims Hy4 outperforms competing open-source models in internal benchmarks. The architecture reflects a deliberate trade-off: by activating only a fraction of total parameters per inference, Tencent reduces compute costs while maintaining model capacity for complex tasks—a key economic lever in an era where inference margins matter as much as training efficiency.
The model targets four verticals where enterprises already embed AI tooling. In software engineering, Hy4 handles code generation, debugging, and validation for multi-file projects. Tencent demonstrated front-end generation (a miniature Three.js scene) from scratch. For game development, the model generates playable Unity prototypes from natural language prompts. In office productivity, it processes multi-document financial audits—filtering, analyzing, and delivering consolidated results. For scientific research, it accelerates molecular dynamics simulations and can iterate improvements autonomously.
Tencent is seeding adoption through two distribution channels. WorkBuddy (enterprise productivity) and CodeBuddy (developer-focused) integrate Hy4 natively across domestic and international versions. Developers can also access the model via Tencent Cloud Tokenhub and third-party platform OpenRouter. WorkBuddy/CodeBuddy users get a two-week free trial.
The open-source release strategy serves dual purposes: it seeds a developer ecosystem that drives adoption of Tencent's proprietary products and cloud services, while positioning the company as a credible open-source contributor in a market where dominance by U.S. labs (OpenAI, Anthropic, Meta) remains unchallenged. Tencent iterated major versions of Hunyuan approximately every two months since rebuilding its infrastructure, releasing preview versions first before full optimization.
