Tencent Hy: Tencent's frontier AI model series | Product Hunt

Hy4 preview

Tencent has unveiled its latest advancement in artificial intelligence, the Hy4 preview model, marking a significant step in its multimodal AI model family. This iteration, following Hy3, represents a substantial scaling up of capabilities, designed to tackle complex, long-horizon agentic tasks. The model's architecture is specifically engineered to handle intricate processes such as autonomous coding, game development, and comprehensive document analysis, indicating a move towards more self-sufficient AI systems. Its design emphasizes not just task execution but also self-correction, as it is built to run its own tests and rectify errors before delivering final outputs.

The development of Hy4 preview involved extensive internal validation, with Tencent leveraging real-world engineering challenges faced by its own specialist teams. This practical approach to training aims to ensure the model's relevance and effectiveness in demanding professional environments. The focus on "long-horizon" tasks suggests an ambition to create AI that can maintain context and coherence over extended periods, a critical feature for complex projects that typically require sustained attention and iterative problem-solving.

770B

The Hy4 preview model boasts an impressive scale, featuring a total of 770 billion parameters, with 49 billion active parameters and a vast 1 million context window. This massive parameter count and extensive context window are crucial for enabling the model's ability to process and understand large volumes of information, which is essential for the complex, multi-step tasks it is designed to perform. The Mixture-of-Experts (MoE) architecture, indicated by the active parameter count, allows the model to efficiently utilize its vast knowledge base by activating only relevant parts for specific tasks, optimizing performance and resource usage.

Internal evaluations conducted by Tencent underscore the model's competitive performance. A blind evaluation involving 163 internal experts across 203 distinct engineering tasks revealed that Hy4 preview slightly outperformed established models like GLM 5.3 and Kimi K3. This comparative success highlights Tencent's growing prowess in the frontier AI space and positions Hy4 as a strong contender among leading large language models, particularly for applications requiring deep technical understanding and problem-solving.

Apache 2.0

Crucially, Tencent has made the Hy4 preview model available under the Apache 2.0 open-source license. This decision to open-source such a powerful and advanced model is a significant contribution to the broader AI community, fostering transparency, collaboration, and accelerated innovation. By providing access to its foundational technology, Tencent enables developers, enterprises, and creators to integrate Hy4 into their own applications, experiment with its capabilities, and build upon its framework without restrictive licensing barriers.

The availability of Hy4 preview through platforms like OpenRouter and WorkBuddy further democratizes access to this cutting-edge AI. This accessibility allows a wider range of users to test, evaluate, and deploy the model in various real-world scenarios, gathering diverse feedback that can inform future improvements. While currently labeled as a "preview," indicating ongoing development and refinement, its open-source nature encourages community engagement in identifying and addressing any remaining quirks, such as its reported tendency to "overthink and over-check itself." This iterative approach, inviting public scrutiny and contribution, aligns with the spirit of open innovation in the rapidly evolving field of artificial intelligence.