Tencent's Hy4 Open-Source Model: 770B Parameters Built for Long-Horizon Agentic Tasks

Tencent Releases 770B Open-Source Model Hy4

Tencent recently launched a large language model preview called Hy4 Preview on Product Hunt, quickly climbing to the #2 spot on the daily leaderboard. The model adopts the mainstream MoE (Mixture of Experts) architecture with a total of 770 billion (770B) parameters, while activating only about 49 billion (49B) parameters per inference — maintaining massive model capacity while significantly reducing inference costs.

What you might have missed is that Hy4 supports a context window of up to 1 million (1M) tokens. This specification means the model can process extremely long documents, entire code repositories, or extended interaction histories in a single pass, providing ample "memory" space for complex tasks. As an open-source model, Hy4's release further strengthens the presence of Chinese-made LLMs in the open-source ecosystem.

Designed for Long-Horizon Agentic Tasks

Hy4's core positioning centers on long-horizon agentic tasks. "Long-horizon" refers to complex tasks that span multiple steps, extend over prolonged periods, and involve interdependent intermediate states. This is fundamentally different from traditional single-turn Q&A conversations.

According to the official description, Hy4 can autonomously handle three typical scenarios:

• Software development: The complete pipeline from requirements understanding to code implementation
• Game development: Compound tasks involving logic, assets, and iteration
• Complex document analysis: Deep comprehension and distillation of massive text volumes

Self-Testing and Self-Repair Closed-Loop Capability

What truly sets Hy4 apart from ordinary code generation models is its claimed ability to "run its own tests and fix bugs before delivery." This means the model doesn't just passively generate code — it can complete an entire work cycle: write code → execute tests → identify issues → autonomously fix them → deliver the final result.

This closed-loop capability represents a key breakthrough direction in the current AI Agent landscape. Traditional models typically end their task after generating seemingly correct code, leaving error detection and fixing to human developers. Hy4 aims to bring this part of the workflow within the model's own capabilities, enabling truly autonomous execution of long-horizon tasks.

Technical Value of MoE Architecture and Ultra-Long Context

From a technical specification standpoint, Hy4's design choices are quite representative. The 770B total parameters / 49B active parameters MoE structure reflects the industry's widespread approach to balancing "expanding model capability" with "controlling computational costs." With active parameters accounting for only about 6.4% of the total, this ultra-large model achieves feasible inference efficiency in real-world deployment.

The 1M token ultra-long context is especially critical for agentic tasks. During long-horizon tasks, the model needs to continuously track previous operation results, code changes, test feedback, and other information. The larger the context window, the lower the probability of the model "forgetting" key information, and the better it can maintain task coherence and consistency. For processing entire code repositories or documents spanning hundreds of pages, sufficient context is a fundamental prerequisite for task completion.

Industry Impact of the Open-Source Strategy

Hy4 is explicitly categorized as an Open Source and Artificial Intelligence project. Tencent's decision to release a model of this scale as open source carries multiple implications.

On one hand, open-sourcing lowers the barrier for developers and enterprises to access cutting-edge model capabilities, fostering a richer application ecosystem around the model. On the other hand, it represents a significant move by Chinese tech companies to establish a stronger voice in the global open-source LLM competition. As multiple Chinese tech firms successively open-source high-performance models, open source has become an important dimension for measuring a company's technical strength and ecosystem influence.

That said, a measured perspective is warranted — Hy4 is still in the Preview stage, and the official capability descriptions lean more toward aspirational positioning. Its actual performance in real-world programming, game development, and document analysis tasks, as well as the reliability of its self-testing and self-repair capabilities, still await broad validation from the community.

Conclusion

The release of Hy4 Preview demonstrates Tencent's continued investment in ultra-large-scale open-source models and the agentic AI direction. The 770B parameter scale, 1M context length, and the positioning of "autonomously completing long-horizon tasks" all point to a clear trend in current AI development — moving from "being able to answer questions" to "being able to get work done."

For developers, if models with self-testing and self-repair closed-loop capabilities can deliver on their promises, they could dramatically reshape how software development and complex task processing are carried out. Whether Hy4 can ultimately prove itself in real-world scenarios will depend on the formal release and more extensive benchmark data to come.