Tencent Releases Hy4, Entering Top Open-Source Model Tier

Original title: 腾讯混元发布 Hy4,首次成为最强开源模型之一
85AI Score

Tencent has released its latest large language model, Hy4, which enters the top tier of open-source models. By scaling up pre-training and post-training, Hy4 introduces visual understanding and Computer Use capabilities for the first time. The model shows significant improvements in instruction following, long-horizon agents, and specialized tasks like financial analysis and 3D construction, with weights to be open-sourced soon under a responsible AI framework.

Full Report

Contents3 sections

Just now, Tencent Hunyuan released its latest-generation large language model, Hy4. With a comprehensive leap in model capabilities, Hy has entered the ranks of the strongest open-source models for the first time. Hy4 will soon have its model weights open-sourced, along with the release of a model system card. We are actively promoting more responsible open-sourcing, with a particular focus on risks related to cybersecurity and biosecurity.

Compared to the Hy4 preview, Hy4 significantly expanded data and training scale during both pre-training and post-training phases. It introduces visual understanding capabilities for the first time and substantially improves overall performance in instruction following, long-context understanding, and long-horizon agent tasks.

Hy4 Benchmark Comparison

Leveraging visual understanding, reasoning, and tool-calling capabilities, Hy4 can perform tasks such as fine-grained image processing, complex chart analysis, and mathematical visual reasoning.

Hy4 also supports Computer Use for the first time, enabling it to combine command-line interfaces and graphical user interfaces to handle various tasks in work and daily life. For specialized software like Blender and CAD, as well as cross-application long-horizon tasks, Hunyuan conducted targeted training, significantly enhancing planning and execution capabilities for professional tasks.

Determining geographic location and verifying information from casual photos

Deep Co-Design Between Models and Products, Continuously Unlocking Real Productivity Scenarios

Thanks to the comprehensive improvement of foundational model capabilities and deep co-design between models and products, Hy4 unlocks more real-world productivity scenarios.

In office scenarios, Hy4 comprehensively enhances the user experience of WorkBuddy. Evaluation results show that Hy4 achieves an average score of 83.18, with a task pass rate of 96.15%. In 9 out of 12 scenarios, it ranks first among domestic models. In actual usage, Hy4 balances effectiveness and efficiency in tasks such as visual creativity, 3D scene construction, and web development. Additionally, for WorkBuddy's unique human-AI collaborative writing feature, Hy4 shows significant improvements in document aesthetics, stability, and usability.

With continuously improving context learning capabilities, the model better understands multi-source complex information, becoming a long-term personal assistant for users. Leveraging WorkBuddy's integration with WeCom (Enterprise WeChat), Hy4 connects daily work information including chat logs, emails, documents, schedules, and meetings. When multiple projects are running in parallel, Hy4 can clarify the progress history and latest updates of each project, identify pending follow-up tasks, meetings requiring preparation, and approaching deadlines, promptly reminding users and assisting with prioritization. Furthermore, team collaboration features powered by Hy4 will launch alongside the new "Project" module in WorkBuddy.

Invoice organization after business trips and financial invoice audits

Hy4 places strong emphasis on enhancing professional financial analysis capabilities. Based on real financial scenarios, thousands of workspaces and tens of thousands of high-quality tasks were constructed, allowing the model to thoroughly learn professional processes and standards in areas such as fund research, securities analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting. Hunyuan collaborated with Tencent Fintech (FiT) to conduct a financial crowdsourced evaluation involving 213 experts, 153 of whom came from top external financial institutions, with 70% having over five years of work experience. Results based on 579 real financial tasks show that Hy4 ranks first among open-source models in overall performance (Crowdsourced scores: Hy4 3.04, Kimi K3.1 Pro 2.68, GLM-5.3 2.68, DeepSeek-V4.1-Flash-1017 2.97). Additionally, in the internal IRAB evaluation by financial AI agent company Xuntu Technology, Hy4 improved by 12.03 percentage points compared to the Hy4 preview.

To enhance practicality in financial and other office scenarios, Hunyuan launched specialized search optimizations targeting timely retrieval, multi-source information integration, and factual consistency verification. In internal tests, Hy4's hallucination rate dropped from 12.95% in the preview version to 7.2%, and the error rate caused by guessing decreased from 23.1% to 5.6%.

In frontend and game development scenarios, Hunyuan performed specialized optimizations for complex frontend interactions, business logic, state management, and front-end/back-end coordination, ensuring that generated interfaces are not only visually appealing but also functionally complete and usable. Hunyuan collaborated with professional game developers at Tencent IEG to continuously optimize data and evaluation quality, focusing on improving capabilities in game functionality implementation, gameplay logic, state understanding, and asset processing. This achieved deep adaptation to mainstream engines such as UE, Unity, and Godot, and began expanding into multiplayer online game development. Compared to the Hy4 preview, Hy4 improved by 19.1% and 9.8% in GameDev and GameLogic evaluations, respectively.

Retro QQ Arcade Hall

In the field of scientific research, Hy4 combines complex reasoning, scientific programming, and long-horizon task execution to further enhance end-to-end capabilities for advancing research work. In the TenureTrail internal evaluation, covering 7 major disciplines and 123 executable research tasks, Hy4 demonstrated solid scientific research capabilities in areas such as mathematical construction, physical modeling, earth environment, systems engineering, and socio-economic decision-making.

Hy4 is actively preparing for the next stage of Recursive Self-Improvement (RSI), enhancing capabilities across various model R&D tasks. Combined with Hyra2.0, Hy4 optimized the pre-training infrastructure for internal small models with the same architecture and automatically trained the Jev model, which outperforms official versions in multiple benchmarks. Meanwhile, Hunyuan has open-sourced the single-cell representation foundation model HyCell and the physical field foundation model HyPDE, developed collaboratively by Hy4 and Hyra2.0, aiming to transform model-assisted research outcomes into reusable tools for the community.

Focus on Cybersecurity and Biosecurity, Continuously Promoting Responsible Open-Sourcing

Addressing potential security risks in cybersecurity and biosecurity, Hunyuan adopts a more responsible approach to training, open-sourcing, and releasing models.

Regarding cybersecurity, Hy4 incorporated only defensive cybersecurity data related to PoC generation and vulnerability patching during training, while excluding offensive training content such as exploit development, reverse engineering, and penetration testing, striving to avoid enhancing potential offensive capabilities. Security experts from Tencent Xuanwu Lab and Keen Labs conducted audits focused on cryptanalysis, reverse engineering, and vulnerability exploitation. The results indicate that Hy4 does not exhibit prominent risks in offensive capabilities such as reverse-solving encryption algorithms, countering binary obfuscation, or vulnerability exploitation.

On the biosecurity front, Hy4’s open-ended experimental design capabilities lag significantly behind those of frontier models (Hy-BioExperimentBench: 30.4/100). An internal audit by the Hunyuan team revealed that while the model demonstrates strong foundational biological knowledge, its ability to translate this knowledge into executable experimental protocols remains limited. No evidence was found suggesting that Hy4 significantly lowers the barrier to conducting biochemical hazardous activities.

Hy4 completes 3D reconstruction of a mouse embryo with 7 million cells in 1.5 hours on a single machine.

Relentless Pursuit of Cost-Effectiveness: Hy4 Flash Coming Soon

Currently, Hy4 is available on Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, Yuanbao, WeCom Smart Assistant Dayuan, and ima. Its API has also been integrated into platforms such as OpenRouter, Hermes, and OpenCode. Integration with other Tencent ecosystem products—including QQ, Honor of Kings, Game for Peace, WeChat Read, and Tencent News—is ongoing.

Guided by the core principle of "model practicality," Tencent Hunyuan strives to push the boundaries of intelligence while providing users with high cost-performance services. Although Hy4 significantly outperforms the Hy4 preview in terms of model performance, pricing remains unchanged: input at ¥6 per million tokens, output at ¥18 per million tokens, and cache hits at ¥0.3 per million tokens.

Hy4 Flash will be released soon. Building upon Hy4’s capabilities, Hy4 Flash drastically reduces task costs, offering users an even more cost-effective model.

To gather more user feedback and continuously enhance model capabilities, WorkBuddy/CodeBuddy will provide a certain quota of free usage from November 4 to November 15. Additionally, from today until December 31, 2026, Hy4 will be billed at 50% off the list price on Tencent Cloud TokenHub.

Since rebuilding its infrastructure in February of this year, the Tencent Hunyuan large model has released a major version approximately every two months, initially opening access via Preview versions before optimizing and releasing official versions based on user feedback.

Simultaneously, Hunyuan continues to scale up both parameter count and data volume. The Hy3 series had fewer than 300 billion parameters, while the Hy4 series reaches 770 billion. It is reported that the next-generation Hy5 will further expand in scale.

The Hunyuan team acknowledges that although Hy4 is currently one of the strongest open-source models, a significant gap remains compared to top-tier closed-source models. Hunyuan will continue to steadily expand pre-training and post-training scales, striving to narrow this gap.

Open in WeChat