Tencent Launches Hy4, One of the Leading Open-Source AI Models

Tencent has unveiled Hy4, its latest-generation open-source large language model, marking a significant advancement in AI performance and capabilities. The new model introduces visual understanding abilities, along with enhancements in instruction following, long-context comprehension, and long-horizon agentic tasks. This enables Hy4 to tackle a wider range of tasks, such as fine-grained image processing, complex chart analysis, and mathematical visual reasoning.

Additionally, Hy4 now possesses computer-use capabilities, allowing it to interact with command-line and graphical user interfaces for various work and everyday tasks.

The model has been trained on professional software like Blender and CAD, which has improved its ability to plan and execute sophisticated tasks. In office productivity applications like Tencent WorkBuddy, Hy4 has shown a notable improvement in performance, achieving an average score of 83.18 and a pass rate of 96.15 percent. It excels in visual creative tasks, 3D scene construction, and web development, while also enhancing document presentation, stability, and usability for collaborative tasks.

Hy4 has been optimized for professional financial analysis, with training on thousands of specialized workspaces and high-quality tasks derived from real-world scenarios. In blind evaluations conducted by Tencent Hy and Tencent Financial Technology, Hy4 emerged as the top-performing open-source model, outperforming other AI models in financial tasks. The model has also reduced its hallucination rate from 12.95% in the Hy4 Preview to 7.2%, and errors caused by unsupported assumptions have decreased from 23.1% to 5.6%.

Additionally, Hy4 has been optimized for frontend and game development, improving performance in complex tasks, business logic, state management, and game logic. The model's collaboration with Tencent Interactive Entertainment Group (IEG) has further enhanced its game development capabilities, including support for major game engines like Unreal Engine, Unity, and Godot.

Hy4 is also advancing into the next stage of Recursive Self-Improvement (RSI) to further enhance its research and development capabilities, with open-sourced foundational models designed for single-cell representations (HyCell) and physics fields (HyPDE).