Tencent has officially launched Hy4, the latest generation of its large language model, marking a significant step forward in the realm of open-source AI. Announced on November 4th, the company states that Hy4’s enhanced capabilities position it among the most potent open models available today. Tencent plans to release the model’s weights and a detailed system card soon, emphasizing its commitment to promoting responsible open access by addressing risks related to network and biological safety.
Hy4: Key Enhancements and New Capabilities
Compared to its predecessor, Hy4 preview, the new Hy4 model boasts substantial improvements in both pre-training and fine-tuning. A standout feature is its integrated visual processing capability, a first for the Hy4 series. Additionally, the model demonstrates significantly enhanced instruction following, better handling of long contexts, improved long-duration task execution, and deeper research functionalities.
Introducing Hy4’s Advanced Features:
- Visual Comprehension: Hy4 is now equipped to process visual information, combining perception, logical reasoning, and tool utilization. This enables it to tackle tasks requiring precise image analysis, complex chart interpretation, and visual mathematical reasoning.
- Computer Interaction Support: The model now supports interaction with computers, fluidly operating through both command-line interfaces and graphical user interfaces.
- Specialized Professional Training: Tencent has specifically trained Hy4 to work with professional software such as Blender and CAD, and to manage long-running inter-program tasks. This specialized training is designed to optimize professional workflows and task execution.
Hy4’s Broad Application Spectrum
With its expanded capabilities, Hy4 is poised for diverse applications aimed at boosting real-world productivity. These include:
- Visual information processing and computer-assisted tasks.
- Front-end development and game creation.
- Financial and office productivity applications.
- Scientific research and model development.
- Long-term personalized assistance.
Performance Benchmarks and Competitive Landscape
Tencent previously released Hy4 preview on August 28th. At that time, the model featured 770 billion general parameters and 49 billion active parameters, with a context length of 1 million tokens. Company representatives noted that Hy4 preview represented a considerable expansion in model size, context length, and data volume, which, coupled with advancements in pre-training and fine-tuning, led to a significant increase in its intelligence level.
Hy4 is currently accessible via Tencent Cloud and OpenRouter, and can be experienced within products like CodeBuddy/WorkBuddy, Yuanbao, and ima. Its performance demonstrably surpasses that of Hy4 preview, while the API pricing remains unchanged.
Looking ahead, Tencent plans to open-source the model’s weights and foster an open ecosystem in collaboration with the vLLM/SGLang and Inferact Inc. communities. A system card detailing the model, including a comprehensive safety assessment, will also be released.
While Hy4 is already recognized as one of the most powerful open-source models available, it still faces a performance gap compared to leading closed-source models. This is particularly evident in challenging benchmarks involving advanced mathematics, long-term development, model iteration, and visual recognition that are developed internally by other entities.
Representatives from Tencent Hunyuan have affirmed the company’s commitment to continuously scaling up pre-training and fine-tuning efforts to bridge this gap and compete with top-tier closed models.
Future Outlook: Hy4 Flash
A forthcoming release, Hy4 Flash, is set to significantly enhance the price-to-performance ratio and reduce inference latency. This version is expected to be integrated more frequently into products, further increasing its practical value.
Source: IThome
