Hy4 – The Flagship Large Language Model from Tencent HunYuan

1. What is Hy4

Hy4 is the official release of Tencent HunYuan's flagship large model. Compared to the preview version, it significantly expands the pre-training and post-training scale, and for the first time introduces visual capabilities and computer operation abilities. It can combine command-line and graphical interfaces to complete professional software operations and long-term tasks. The model performs exceptionally well in real productivity scenarios such as financial research, game development, front-end generation, and scientific computing. Official evaluations place it in the first tier of open-source models. It is now available on Tencent Cloud and OpenRouter, and can be experienced in products such as CodeBuddy, WorkBuddy, Yuanbao, and ima. The API pricing remains the same as the preview version.

Technical positioning and domain: Hy4 belongs to the field of natural language processing and multi-modal large models, and is positioned as a general-purpose flagship model for real productivity scenarios. Its technical approach continues the Mixture of Experts (MoE) architecture from the preview version, but significantly expands the pre-training and post-training scale. For the first time, it integrates visual encoding capabilities into a unified model, achieving deep coupling between the text reasoning chain and image perception results. The model supports a context window of up to one million tokens, enabling long-text understanding, in-depth research, and cross-application long-term task execution, establishing a differentiated technical positioning among similar open-source models in the industry.

Development background: Hy4 was developed by the Tencent HunYuan team, leveraging Tencent's vast real business data and technical expertise in areas such as social media, gaming, finance, and office productivity. The development motivation clearly targets solving the challenges of deploying large models in real productivity scenarios, i.e., transitioning from benchmark capabilities to actual business effectiveness. Tencent internally collaborates with teams such as IEG (Interactive Entertainment Group) and the financial business line to build a large-scale reinforcement learning and preference alignment signal source based on real business scenarios, making the model's capabilities more aligned with practical application needs. In addition, the co-released Hyra 2.0 scientific agent, along with HyCell and HyPDE foundational models for scientific research, further strengthens Tencent HunYuan's technical layout in the AI for Science domain.

Core value: Hy4 addresses capability gaps in complex scenarios such as professional software operation, cross-application long-term tasks, and multi-modal understanding. Through specialized post-training, the model learns a cycle of observing, planning, calling tools, and verifying results, enabling it to operate professional software such as Blender and CAD. In financial scenarios, the hallucination rate drops from 15.76% to 5.45%, and the task completion rate for office document processing approaches 97%. These figures indicate that Hy4 has evolved from a simple text generation tool into an autonomous agent capable of planning and executing tasks, providing practical AI solutions for professional fields such as financial research, game development, and scientific computing.

2. Key Features

• Visual Understanding and Computer Interaction: Hy4 introduces visual encoding capabilities for the first time, integrating image perception results directly into the text reasoning chain. This supports fine-grained image processing, complex chart analysis, and mathematical visual reasoning. At the same time, the model can flexibly combine command-line and graphical interface operations, just like a human, to complete long-range cross-application tasks. This capability is highly technically scarce among open-source models.

• Frontend and Game Development: The model has been specially optimized for complex interactions, business logic, state management, and frontend-backend coordination. It is deeply compatible with mainstream game engines such as UE, Unity, and Godot. According to official evaluation data, GameDev and GameLogic benchmarks improved by 19.1% and 9.8%, respectively, compared to the preview version, demonstrating that the joint training strategy of visual understanding and frontend code generation indeed delivers measurable performance gains.

• Finance and Office Work: The model systematically learned professional processes such as fund research, securities analysis, and quantitative strategies. In the official public testing of 579 real tasks, it scored 3.04/4, ranking first among open-source models. The hallucination rate dropped from 15.76% to 5.45%, and it supports real-time collaborative editing of Word, Excel, and PPT documents with humans, covering specific office workflows such as invoice organization and tax audits.

• Scientific Research and Model Development: Combining complex reasoning, scientific programming, and long-range task execution, Hy4 can advance scientific research in an end-to-end manner. With the help of the Hyra 2.0 agent, the model can automatically train models that outperform the official Jev model. It has also open-sourced two foundational research models, HyCell and HyPDE. In the TenureTrail benchmark (7 disciplines, 123 tasks), it also ranked first among open-source models.

• Long-Term Personal Assistant: Context learning capabilities have been significantly enhanced. Through the WorkBuddy WeCom enterprise suite, it connects chat records, emails, documents, schedules, and meetings, automatically identifying to-do items, reminding of schedules, and assisting in prioritizing tasks. A 1M token context window provides the technical foundation for long-term memory and multi-project parallel management.

• Agentic Tool Calling and Specialized Post-Training: The model has undergone specialized training for professional software (Blender, CAD) and long-range cross-application tasks, learning the complete cycle of observation—planning—tool calling—result verification. In the financial scenario, the training environment was built based on thousands of workspaces and tens of thousands of high-quality tasks, with specialized search optimizations for time-sensitive retrieval, multi-source information integration, and factual consistency checks.

3. Application Scenarios

• Financial Research and Investment: Analysts can use Hy4 to complete fund research, securities analysis, quantitative strategies, valuation modeling, and individual stock research report writing. In the official public testing, this scenario ranked first among open-source models, with a significant reduction in hallucination rate. It also supports daily office processes such as invoice organization and financial audit checks, making it suitable for use by securities firms, fund companies, and independent investment researchers.

• Game Development: Developers can leverage Hy4 within mainstream engines like UE, Unity, and Godot to implement functional development, gameplay logic writing, and asset processing. This capability has been extended to multiplayer online game development. Tencent has already developed an actual playable project called "QQ Happy Arcade Hall" based on Hy4, verifying the model's feasibility in generating complex game logic.

• Frontend and 3D Development: Frontend engineers can use Hy4 to generate complex interactive frontend interfaces with complete business logic, as well as works such as Three.js 3D digital sandboxes. The model's joint training advantages in visual understanding and frontend code generation enable it to produce interfaces that are both aesthetically pleasing and user-friendly, making it ideal for rapid prototyping and interaction design validation.

• Scientific Research and Computation: Researchers can use Hy4's end-to-end capabilities to advance their scientific work. Officially demonstrated cases include completing the 3D reconstruction of a 7 million-cell mouse embryo in 1.5 hours, discovering counterexamples for mathematical conjectures, and performing Lean formal verification. Combined with the Hyra 2.0 agent, the model can automatically complete the entire scientific research process, from experimental design to result analysis.

• Personal Office and Long-term Assistant: Through the WorkBuddy WeCom enterprise suite, Hy4 can integrate chat records, emails, schedules, and meeting information to automatically generate daily task dashboards, reminders, and assist in prioritizing tasks. It also supports dual-editing between human and machine for Word, Excel, and PPT documents, making it suitable for professionals who require efficient time management and document collaboration.

4. How to Use

1. Choose an Entry Point: Regular users can directly open Tencent products such as Yuanbao, ima, WorkBuddy, or CodeBuddy to experience Hy4, without requiring any additional environment configuration. Developers can call the API via Tencent Cloud or OpenRouter to integrate the model into their own applications or workflows.

2. Obtain Permissions: End-user products can be used by logging into a Tencent account. API users need to first activate the Hunyuan service on Tencent Cloud and obtain an API Key, then configure the model name as Hy4 in their code to initiate calls.

3. Call the Model: Developers set the model name to Hy4 in their code and configure the corresponding API endpoint and authentication information. Both Tencent Cloud and OpenRouter provide standard RESTful API interfaces, supporting streaming output and batch requests, making it easy to integrate into various applications.

4. Integrate into Scenarios: Connect to the relevant capabilities based on your needs. For example, use WorkBuddy to integrate with WeCom for long-term assistance, automatically consolidating chat history, emails, and schedules; or perform multimodal Q&A, document co-writing, and in-depth research directly within Yuanbao or ima. API users can select different capability modules such as visual understanding, text generation, or tool calling based on their business scenarios.

5. Experience Advanced Capabilities: To use advanced features such as visual understanding, computer operation, game development, or scientific computing, submit tasks via Agent in Hy AI Studio or CodeBuddy, allowing Hy4 to autonomously plan and execute them. It is recommended to start with simple tasks on the first use, gradually becoming familiar with the model's instruction understanding and tool calling patterns.

5. Pros and Cons Analysis

6. Comparative Analysis with Similar Tools

Selection Recommendations: For professional productivity scenarios such as financial research, game development, and scientific computing, Hy4 demonstrates a clear advantage with its specialized capabilities trained on real business data and its differentiated features in visual and computer interaction, especially for enterprise users already within the Tencent ecosystem. If cost-effective general-purpose dialogue and code generation are the priorities, DeepSeek V3 offers a more flexible deployment option with its MIT open-source license and lower pricing. Kimi K3.1 Pro maintains its traditional strengths in long-text processing and general-purpose dialogue, making it suitable for users whose core needs revolve around text processing. Overall, the three models have clearly distinct positioning, and selection should prioritize alignment with one's own business scenarios and ecosystem.

7. Editor's Summary

The release of Tencent HunYuan Hy4 marks a significant step forward in the practical productivity capabilities of domestic open-source large models. From a technological innovation perspective, Hy4 is the first to integrate visual encoding capabilities into a unified model and supports dual-modal computer operations with GUI and command-line interfaces. This technical approach is relatively rare among open-source models. More importantly, its training strategy has shifted—from relying solely on public benchmarks to deeply integrating real business data from Tencent's internal financial, gaming, and office operations to build reinforcement learning signal sources. This enables the model's capabilities to transition from laboratory metrics to actual business performance. Data such as the hallucination rate in financial scenarios dropping from 15.76% to 5.45% and a 19.1% improvement in GameDev evaluations directly demonstrate the effectiveness of this training strategy.

From a practical value standpoint, Hy4 has already demonstrated the early stages of a real productivity tool in professional fields such as financial research, game development, front-end generation, and scientific computing. Its 1M token context window provides the technical foundation for long-text understanding and long-term personal assistant capabilities. Meanwhile, the integration with the WorkBuddy enterprise WeChat suite opens up a clear path for deployment in office scenarios. For professional users such as financial analysts, game developers, and researchers, Hy4 offers end-to-end AI assistance from task planning to execution, a feature that is not commonly found in current open-source models.

The primary target users of Hy4 fall into three categories: first, enterprise users within the Tencent ecosystem who can seamlessly integrate with products such as WorkBuddy and Yuanbao; second, developers in specialized fields such as finance, gaming, and scientific research who can build vertical applications using APIs or Agent modes; and third, AI application developers requiring multi-modal understanding and computer operation capabilities. Its future development potential hinges on two key factors: the actual quality of the open-source weights released and the construction of the community ecosystem, as well as the stability validation of its visual and computer operation capabilities in more real-world scenarios. If these two aspects continue to improve, Hy4 has the potential to establish a differentiated competitive advantage within the open-source large model landscape.

8. FAQ

Q: What are the main differences between Hy4 and the preview version?
A: Compared to the preview version, the official release of Hy4 significantly expands the pre-training and post-training scale, and introduces visual encoding capabilities and computer operation capabilities for the first time. Frontend and game development capabilities have seen clear improvements (GameDev +19.1%, GameLogic +9.8%), the hallucination rate in financial scenarios has dropped from 15.76% to 5.45%, and the overall capability has evolved from pure text generation to a multi-modal understanding and tool-calling agent form.

Q: What is the price of Hy4's API? How to integrate it?
A: The API pricing for Hy4 is the same as the preview version: 6 RMB per 1 million input tokens, 18 RMB per 1 million output tokens, and 0.3 RMB per 1 million cached tokens. Developers can access the Hy4 API by enabling the HunYuan service on Tencent Cloud to obtain an API Key, or by directly calling it on the OpenRouter platform. Simply configure the model name as "Hy4" in the code.

Q: Is Hy4 open source? Can it be deployed locally?
A: The official plan is to open source Hy4's weights and release system cards, but as of the time of this article's publication, the specific release date and hosting channels for the official weights have not been announced. Considering that the preview version already has a total parameter count of 770B and 49B active parameters, local deployment would require significant hardware investment. It is recommended to follow the official Tencent HunYuan channels for the latest open-source updates.

Q: How can users utilize Hy4's visual and computer operation capabilities?
A: Regular users can directly experience multi-modal Q&A and image understanding through products like Yuanbao and ima. Developers can submit tasks via Hy AI Studio or CodeBuddy in an Agent format, allowing Hy4 to autonomously plan and execute computer operation workflows. This capability supports operating professional software such as Blender and CAD, and can complete long-range tasks across multiple applications.

Q: How does Hy4 perform in financial scenarios?
A: In the official crowd testing, Hy4 completed 579 real financial tasks, evaluated by 213 experts, achieving a score of 3.04/4, ranking first among open-source models. The model has systematically learned professional processes such as fund research, securities analysis, and quantitative strategies. The hallucination rate has dropped from 15.76% to 5.45%, and the system has undergone specialized optimization for time-sensitive retrieval, multi-source information integration, and factual consistency verification.

Q: Compared to Kimi K3.1 Pro, which one is stronger?
A: The two models are positioned differently. Hy4 performs better in professional productivity scenarios such as financial crowd testing (3.04/4 vs 2.68/4), scientific research evaluations, and game development, and also features visual and computer operation capabilities. Kimi K3.1 Pro, on the other hand, maintains its traditional advantages in long-text processing and general conversation. The choice should be based on the specific business scenario.

9. Project Links

• Product Official Website: https://hy.tencent.com/research/hy4
• Tencent Yuanbao (consumer experience entry): https://yuanbao.tencent.com
• Tencent IMA (knowledge management experience entry): https://ima.qq.com
• Tencent Cloud Hunyuan Service (API integration): https://cloud.tencent.com/product/hunyuan
• OpenRouter (third-party API calling): https://openrouter.ai