Tencent Hybrid Officially Releases New Generation Big Language Model Hy4 Introduces Visual Capabilities for the First Time

The Zhitong Finance App learned that on November 4, Tencent Hybrid will release the official version of Hy4. With another full jump in model power, Hy became one of the strongest open source models for the first time. Compared to the Hy4 preview, Hy4 significantly expanded the scale of pre-training and post-training, introduced visual ability for the first time, and greatly improved comprehensive abilities such as following instructions, understanding long texts, long-range tasks, and in-depth research.

Hy4 introduced visual abilities for the first time. It can combine perception, reasoning, and tool call to perform tasks such as fine-grained image processing, complex chart analysis, and mathematical visual reasoning. Hy4 also supports computer operation for the first time, and can be flexibly combined with a command line and graphical interface. Tencent Hybrid has conducted special training for professional software such as Blender and CAD and cross-application long-range tasks to continuously optimize the planning and execution of professional tasks.

The introduction of visual understanding has also significantly improved Hy4's front-end and game development capabilities. Tencent Hybrid has targeted optimization for complex front-end interaction, business logic, state management, front-end and back-end linkage, etc., so that the model-generated interface is not only beautiful, but also fully functional and usable. In the game development scenario, Tencent Hybrid collaborated with Tencent IEG professional game developers to continuously optimize data and evaluation quality, focusing on improving game function implementation, gameplay logic, state understanding, and asset processing capabilities, achieving deep adaptation to mainstream engines such as UE, Unity, and Godot, and beginning to expand to multiplayer online game development. Compared to the Hy4 preview, Hy4 improved 19.1% and 9.8% in GameDev and GameLogic reviews, respectively.

Tencent's mixed elements have built thousands of workspaces and tens of thousands of high-quality tasks based on real financial scenarios, so that the model can fully learn professional processes and specifications in the fields of fund research, securities analysis, quantitative strategies, asset allocation, corporate finance, tax accounting, etc. Tencent Hybrid and Tencent Fintech FiT conducted a financial public survey of 213 experts, of which 153 were from leading external financial institutions, and 70% had more than five years of work experience. Based on 579 real financial tasks, Hy4's overall performance ranked first in the open source model (public test score: Hy4 3.04, Kimi K3.1 Pro 2.68, GLM-5.3 2.68, DeepSeek-V4.1-Flash-1017 2.97). Furthermore, in the internal evaluation IRAB of the financial intelligence company XunTu Technology, Hy4 increased 12.03 percentage points compared to Hy4 preview.

In order to improve the practicality of various office scenarios such as finance, Tencent Hybrid has also initiated special search optimization for time-sensitive search, multi-source information integration, and factual consistency verification. In self-built tests, Hy4's illusion rate dropped from 15.76% of the preview to 5.45%, and the guess error rate dropped from 23.1% to 5.6%. In addition, Hy4 also focuses on improving human-computer dual-writing capabilities in office scenarios, and supports users to collaborate with Hy to edit office files such as Word, Excel, and PPT in real time. The blind review showed that Hy4 compared to Hy4 preview had significant improvements in document aesthetics, stability, and usability.

Hy4 combines complex reasoning, scientific programming, and long-range task execution to further enhance the ability to advance scientific research work from end to end. In TenureTrail's internal evaluation, which covered 7 university disciplines and 123 research tasks, Hy4 showed solid scientific research capabilities in fields such as mathematical construction, physical modeling, global environment, systems engineering, and socioeconomic decision-making, ranking first among open source models.

Hy4 is also preparing for recursive self-improvement (RSI) to improve capabilities for various model development tasks. Combined with the Hyra2.0 scaffolding, Hy4 optimized the internal pre-training infrastructure for small models with the same architecture, and automatically trained to be superior to the official Jev model in many evaluations. At the same time, Tencent Hybrid also open-sourced HyCell, a basic model for single-cell characterization developed by Hyra2.0 in collaboration with Hyra2.0, and HyPDE, a basic physics-based model, to transform the results of model-assisted research into tools that can be reused by the community.