Hy4 Operational

Hy4 is the official release of Tencent Hunyuan's flagship foundation model. Compared with the preview version, it significantly expands pre-training and post-training scale, and introduces vision and computer-use capabilities for the first time. It can combine command-line and graphical interfaces to complete professional software and long-horizon tasks. Hy4 performs strongly in real-world productivity scenarios including financial research, game development, frontend generation, and scientific computing. According to official evaluations, it ranks among the top tier of open-source models. It is now available on Tencent Cloud and OpenRouter, and can be experienced through products such as CodeBuddy, WorkBuddy, Yuanbao, and ima. Its pricing is the same as the preview version.

Visual understanding and computer use: Hy4 introduces vision capabilities for the first time, combining perception, reasoning, and tool use for fine-grained image processing, complex chart analysis, and mathematical visual reasoning. It also supports flexible, human-like computer operation through a combination of command-line and graphical interfaces for the first time.
Frontend and game development: Optimized for complex interactions, business logic, state management, and frontend-backend coordination, Hy4 is deeply adapted to mainstream engines such as UE, Unity, and Godot. Its GameDev and GameLogic evaluation scores improved by 19.1% and 9.8%, respectively, over preview.
Finance and office work: Trained on professional workflows including fund research, securities analysis, and quantitative strategies. It ranked first among open-source models in official public testing, reducing the hallucination rate from 15.76% to 5.45%. It also supports real-time collaborative editing of Word, Excel, and PPT documents.
Scientific research and model development: By combining complex reasoning, scientific programming, and long-horizon task execution, Hy4 can advance research work end to end. With Hyra 2.0, it automatically trained a model superior to the official Jev and open-sourced two foundational research models, HyCell and HyPDE.
Long-term personal assistant: Improved in-context learning enables the WorkBuddy WeCom suite to connect chat records, email, documents, calendars, and meetings. It can automatically identify to-dos, remind users of schedules, and help prioritize tasks.

Scale expansion and sparse activation: Compared with preview, Hy4 significantly expands pre-training and post-training scale while continuing preview's mixture-of-experts approach. Preview has 770B total parameters and 49B active parameters. Hy4 also maintains a 1M-token context window to support long-document understanding and long-horizon tasks.
Multimodal fusion: Hy4 incorporates visual encoding into a unified model for the first time, allowing its text reasoning chain to call on image perception results for fine-grained image processing, chart analysis, and visual mathematical reasoning. Improvements in frontend and game development also result from joint training across visual understanding, code generation, and engine tool use.
Agentic tool use and specialized post-training: Hy4 receives specialized training for professional software such as Blender and CAD, as well as cross-application long-horizon tasks. It learns the cycle of observing, planning, calling tools, and verifying results. For financial scenarios, training environments were built from thousands of workspaces and tens of thousands of high-quality tasks, with specialized search optimization for time-sensitive retrieval, multi-source information integration, and factual consistency checks to reduce hallucinations and speculative errors.
RL-driven alignment with real-world tasks: Real business scenarios including games, in collaboration with IEG, finance, based on feedback from 213 experts, office documents, and Honor of Kings configuration and bug analysis provide signals for large-scale reinforcement learning and preference alignment. This helps transfer model capabilities from benchmarks to real-world productivity.

Choose an entry point: General users can open Yuanbao, ima, WorkBuddy, CodeBuddy, or other products to experience Hy4 directly. Developers can access the API through Tencent Cloud or OpenRouter.
Obtain access: Users of consumer products can sign in with a relevant Tencent account. API users must enable the Hunyuan service on Tencent Cloud and obtain an API key.
Call the model: Configure the model name as Hy4 in your code.
Integrate with your workflow: Connect the required capabilities for your use case, such as using WorkBuddy as a long-term assistant connected to WeCom, or conducting multimodal Q&A, dual-entry document editing, and deep research directly in Yuanbao or ima.
Experience advanced capabilities: For vision, computer use, game development, or scientific computing, submit tasks as an agent in Hy AI Studio or CodeBuddy and let Hy4 plan and execute them autonomously.

Top-tier open-source performance: Hy4 scored 3.04/4 on 579 real-world financial testing tasks, ranking first among open-source models. It also ranked first among open-source models in the scientific TenureTrail evaluation. Officially, it is described as having “joined the ranks of the strongest open-source models for the first time.”
Vision and computer use introduced for the first time: Hy4 adds multimodal perception and GUI-plus-command-line computer operation, enabling it to use professional software such as Blender and CAD and complete cross-application long-horizon tasks.
Deeply refined for real-world productivity: The hallucination rate fell from 15.76% to 5.45%, while office document task completion approached 97%. Specialized training based on real Tencent business data supports frontend and game development, including a 19.1% GameDev improvement, as well as financial research scenarios.
Million-token context and long-horizon tasks: The 1M-token context window is maintained, significantly improving long-document understanding, deep research, and long-term assistant capabilities across multiple concurrent projects.

Official website: https://hy.tencent.com/research/hy4

Financial research: Conduct fund research, securities analysis, quantitative strategy development, valuation modeling, and individual stock research, while ranking first among open-source models in financial testing. It also supports office workflows such as invoice organization and tax inspection.
Game development: Deeply adapted to mainstream engines including UE, Unity, and Godot, covering feature implementation, gameplay logic, and asset processing. It has expanded to multiplayer game development, such as the actually playable “QQ Happy Arcade.”
Frontend and 3D development: Generate frontend interfaces with complex interactions and complete business logic, as well as projects such as Three.js 3D digital sandboxes, while balancing aesthetics and usability.
Scientific computing: Advance research work end to end, such as completing the 3D reconstruction of a 7-million-cell mouse embryo in 1.5 hours, discovering counterexamples to mathematical conjectures and performing Lean formal verification, and autonomously optimizing model-training infrastructure.
Personal office work and long-term assistance: The WorkBuddy WeCom suite integrates chat records, email, calendars, and meetings to automatically generate daily task boards, remind users of to-dos, and help prioritize tasks. It also supports human-AI collaborative editing of Word/Excel/PPT.