正文 · AI 翻译
Thanks to IT Home readers 栀夏, Domado, McWings, xxy171070, Alita的迷弟, 吃瓜群众_, 八重神子, and 未央 for the tips!
IT Home reported on November 4 that Tencent Hunyuan released the official version of Hy4, its latest generation large language model, today (the 4th). According to the company, another broad leap in model capabilities has put Hy among the strongest open-source models for the first time. Tencent plans to open-source Hy4’s model weights soon, publish a model system card, and actively pursue more responsible open sourcing, with particular attention to risks including cybersecurity and biosecurity.
Compared with Hy4 preview, Hy4 has significantly expanded the scale of pretraining and post-training, introduced visual capabilities for the first time, and substantially improved its overall capabilities in instruction following, long-text comprehension, long-horizon tasks, and deep research.
Hy4 introduces visual capabilities for the first time, combining perception, reasoning, and tool use to perform tasks such as fine-grained image processing, complex chart analysis, and mathematical visual reasoning. Hy4 also supports computer use for the first time and can flexibly combine command-line and graphical interfaces. Tencent has also conducted specialized training for professional software such as Blender and CAD, as well as long-horizon tasks spanning multiple applications, to continue improving its ability to plan and execute professional tasks.
According to the introduction, the model can be used in a range of scenarios, including visual and computer tasks, frontend and game development, finance and office work, scientific research and model development, and long-term personal assistance. It is designed for real-world productivity.
As IT Home previously reported, Tencent released Hy4 preview on August 28, with 770B total parameters, 49B active parameters, and a 1M context length.
At the time, the company said Hy4 preview had been significantly expanded in model size, context length, and data scale, and that advances in both pretraining and post-training had brought another major improvement in intelligence.
Hy4 is now available on Tencent Cloud and OpenRouter, and can be tried in products including CodeBuddy / WorkBuddy, Yuanbao, and ima. It significantly outperforms Hy4 preview while retaining the same API price.
The company plans to open-source the model weights and has partnered with the vLLM/SGLang communities and Inferact Inc. to support the open-source ecosystem. It also plans to publish a model system card at that time, including details such as more comprehensive safety evaluations.
Although Hy4 is already one of the strongest open-source models available, a considerable gap remains between it and leading closed-source models, especially on difficult internal test sets covering frontier mathematics, long-horizon development, model iteration, and visual question answering. Tencent Hunyuan said it will continue to steadily expand the scale of pretraining and post-training in an effort to narrow the gap with leading closed-source models.
Hy4 flash is also planned for release soon. It is expected to significantly improve the model’s cost-effectiveness and inference speed, and to be updated more frequently in coordination with products to continually improve its practical usefulness.