Hy4 preview - Tencent's next-generation flagship large-scale model from Hunyuan Open Source

What is Hy4 preview?

Hy4 preview is Tencent's flagship open-source model, boasting 770 bytes of parameters, 49 bytes of activation parameters, and a 1MB context length. Developed in collaboration with experts from Tencent's software engineering, gaming, finance, and security sectors, the model demonstrates exceptional performance in real-world productivity tasks such as code development, office analytics, game creation, and scientific research. It consistently ranks among the top open-source models in multiple authoritative benchmark tests and blind tests across 203 engineering tasks. Furthermore, it has achieved substantial breakthroughs in scientific research, including molecular dynamics simulations, quantum device design, and addressing long-standing geometric challenges.

Main functions of Hy4 preview

• Software EngineeringEnhance the ability to understand, plan, debug, and verify long-term development tasks, and further improve the visual aesthetics and interaction quality of front-end development.
• Office AnalysisSignificantly enhances the ability to understand complex office environments and perform financial analysis, enabling the completion of the entire process from information processing to the delivery of documents, spreadsheets, and presentations.
• Game developmentEnhance the ability to generate playable prototypes directly from a single requirement, master game engines, and support multi-round interactions to continuously improve complex game projects.
• Scientific researchIt significantly enhances the ability to understand, reason, and solve complex scientific research problems, and has made great strides in AI research and development, molecular dynamics simulation, condensed matter physics, and basic mathematics.
• Inference optimizationThe system autonomously analyzes and infers system bottlenecks, and conducts multiple rounds of optimization around operator fusion and communication optimization, resulting in a 31.8% improvement in end-to-end throughput compared to the baseline.
• Experimental CoordinationIt manages multiple Codex Sessions to conduct experiments like a researcher, outperforming Codex Independent Exploration in all 8 metrics, demonstrating the ability to determine direction, organize experiments, and continuously iterate in complex R&D tasks.

Hy4 preview's technical principles

• Software EngineeringHy4 preview significantly enhances the ability to understand, plan, debug, and verify long-term development tasks. In the front-end development of large projects, it can proactively improve visual aesthetics and interaction quality, and the generated code can be directly run and continuously iterated in a real engineering environment.
• Office AnalysisThe model significantly enhances its understanding of complex office environments through high-quality data co-construction with financial experts. It can handle tasks such as cross-document collaboration and data analysis, and complete the entire office workflow from raw information processing to final document, spreadsheet and presentation delivery, such as automatically reviewing 24 expense reports and cross-referencing policy documents.
• Game developmentHy4 preview has the ability to generate playable prototypes with just one sentence and is proficient in using mainstream game engines such as Unreal Engine 5 and Unity. Developers can continuously improve complex game projects through multiple rounds of natural language interaction, covering the complete development chain, including scene building, gameplay logic, character animation blending tree, and sound effects system.
• Scientific researchIn scientific research scenarios, the model demonstrates a deep understanding, reasoning, and problem-solving ability for complex problems. When used in conjunction with the Hyra agent, it accelerates the simulation of molecular dynamics in machine learning force fields by 2.0 times, reduces the leakage rate from 48.2% to 4.8% in the design of low-temperature quantum transport devices, and pushes the lower bound of volume to 0.41104 in the century-old classical geometric problem 3D Blaschke-Lebesgue.
• Inference optimizationHy4 preview can autonomously analyze the performance bottlenecks of the inference system and carry out multiple rounds of iterative optimization in areas such as operator fusion and communication optimization. The end-to-end service throughput is improved by 31.8% compared to the baseline. Stable benefits are achieved under different context lengths and concurrency levels, verifying the model's ability to autonomously locate bottlenecks and continuously optimize the infrastructure.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Hy4 preview

• Open source downloadGet model weights and code via HuggingFace, GitHub, ModelScope, or GitCode, supporting local deployment and secondary development.
• API callsConnect to Tencent Cloud TokenHub or OpenRouter, and directly call the model interface with a pricing of 6 yuan per million tokens as input and 18 yuan per million tokens as output.
• Product ExperienceExperience Hy4 preview's office and programming capabilities directly in Tencent's products such as WorkBuddy, CodeBuddy, Yuanbao, and ima.
• Online trial playVisit the online model experience website https://aistudio.tencent.com/ to interact with the model online without configuring any environment.

Hy4 preview's core advantages

• Leading in scaleWith a total of 770B parameters, 49B activation parameters, and a context length of 1M, it firmly ranks among the top open-source models in terms of model size and long text processing capabilities.
• Real productivity orientationIt collaborates deeply with top experts in software engineering, gaming, finance, and security within Tencent, focusing on real-world scenarios such as code, office work, and gaming, rather than simply manipulating rankings.
• Scientific research breakthrough capabilityIn conjunction with Hyra, it accelerates molecular dynamics simulations by 2.0 times, reduces the leakage rate of quantum devices to 4.8%, and pushes the lower volume bound of the century-old geometric problem 3D Blaschke-Lebesgue to 0.41104.
• Self-optimizing reasoningIt can independently analyze system bottlenecks and continuously iterate around operator fusion and communication optimization, resulting in a 31.8% improvement in end-to-end service throughput compared to the baseline.
• Intelligent Experiment CoordinationManage multiple Codex Sessions in parallel like a researcher, outperforming Codex standalone exploration in all 8 evaluation metrics in the small model post-training task.
• Expert blind test leadsIn blind tests conducted by 163 internal experts across 203 engineering tasks, Hy4 preview achieved an average score of 2.99/4.00, slightly outperforming GLM-5.3 and Kimi K3.

Hy4 preview project address

• Project official website：https://hy.tencent.com/research/hy4-preview
• GitHub repository：https://github.com/Tencent-Hunyuan/Hy4-preview
• HuggingFace model library：https://huggingface.co/tencent/Hy4-preview

Hy4 preview's comparison with similar competing products

Comparison Dimensions	Hy4 preview	DeepSeek-V4 Pro
Total parameters / Activation parameters	770B / 49B	Approximately 671B / 37B
Context length	1M	128K
Model Architecture	MoE (Hybrid Expert)	MoE (Hybrid Expert)
Open source status	Fully open source (weight + code)	Open source weights, open source inference code
Core positioning	Real productivity (coding, office work, gaming, scientific research)	General Reasoning and Code
Research capabilities	Hyra's breakthroughs in molecular dynamics, quantum devices, and mathematical proofs	Strong general reasoning; research toolchain needs to be built independently.
Product Collaboration	Deep integration with WorkBuddy, CodeBuddy, Yuanbao, and IMA	Primarily relies on third-party integration and self-deployment
API pricing (input/output)	6 yuan / 18 yuan (per million tokens)	Approximately 2 yuan / 8 yuan (per million tokens)
Internal blind test performance	With an average score of 2.99/4.00, it is slightly better than GLM-5.3 and Kimi K3.	It performed well in benchmark tests, but there are no publicly available blind tests of similar engineering applications.

Application scenarios of Hy4 preview

• Intelligent code developmentIt supports long-term project understanding, front-end visual aesthetic optimization, debugging and verification, and can be used with CodeBuddy to complete the entire development process from requirements to runnable code.
• Complex office automationIt can handle cross-document data analysis, financial audits, and system comparisons, and automatically generate documents, tables, and presentations. It has been implemented in office products such as WorkBuddy.
• Rapid Build of Game PrototypesA playable prototype can be generated from a single natural language requirement. Proficient in using the Unity/Unreal 5 engine, it supports multi-round interactions to continuously improve levels, characters, and gameplay.
• Accelerated scientific researchIt works in conjunction with the Hyra agent to conduct molecular dynamics simulations, quantum device design, and mathematical theorem proofs, autonomously managing the experimental process and continuously iterating and optimizing the direction.
• 3D digital content generationGenerate interactive Three.js websites and digital twin scenarios based on precise parameters, supporting GLB model loading, procedural modeling, and multi-language responsive deployment.