Tencent has officially released the full version of Hy4, joining the ranks of the strongest open-source models. Compared with the Hy4 preview, this version expands the scale of both pre-training and post-training, introduces visual capabilities for the first time, and significantly improves comprehensive capabilities such as instruction following and long-document understanding—unlocking real productivity across multiple scenarios. In the areas of vision and computer tasks, Hy4 can handle fine-grained image processing and complex chart analysis, supports computer operation, and has been specially optimized for professional software and long-horizon cross-application tasks. Front-end and game development capabilities have been substantially enhanced, achieving deep adaptation to multiple mainstream game engines, with related benchmark results greatly improved. In financial scenarios, a large number of tasks were built based on real-world situations and joint expert crowdsourced testing was conducted, with comprehensive performance ranking first among open-source models. At the same time, search-specific optimization has been completed, greatly reducing hallucination rates and guess error rates, and it also supports multi-user real-time collaborative editing of various office documents. In the fields of scientific research and model development, Hy4 ranks first among open-source models in relevant internal evaluations, can autonomously complete complex scientific research tasks, and is preparing for recursive self-improvement—having already collaboratively developed multiple open-source foundation models with excellent performance. In addition, Hy4's in-context learning capability has improved, allowing it to serve as a long-term personal assistant connected to various daily work information streams, helping users sort out follow-up items and prioritize tasks. The team will continue to expand the scale of model training, narrow the gap with top closed-source models, and plans to release Hy4 Flash—a cost-effective model with high inference speed—in the near future.