Tencent’s new Hy4 preview model is pushing AI-assisted software development into a more agentic direction, with the company demonstrating how a single natural-language prompt can be turned into a playable game prototype. The open-weight model can work with game engines and continue refining projects through multiple interactions, potentially reducing the amount of manual coding required to move from an idea to a working application.
Released on August 28, 2026, Hy4 preview is Tencent’s latest Hunyuan model and is designed for software engineering, office productivity, game development and scientific research. It has 770 billion total parameters, 49 billion active parameters and a context window of up to 1 million tokens. Tencent has released the model under the Apache License 2.0, making its weights available for developers and organizations to use and adapt.
Tencent Hy4 Can Build Games From A Single Prompt
One of the most notable capabilities demonstrated by Tencent is Hy4 preview’s ability to turn a natural-language request into a playable game prototype.
The company says the model can work with game engines and allow developers to continue refining complex projects through multiple turns.
This means a developer can describe the desired gameplay and have Hy4 handle parts of the development process rather than simply returning snippets of source code.
Hy4 Game Development Workflow
Natural-language prompt
│
▼
Hy4 preview
│
├── Understand requirements
├── Plan project
├── Generate code
├── Work with game engine
├── Test output
└── Refine across turns
│
▼
Playable prototype
The approach is particularly significant for game development because building even a simple playable experience normally requires multiple disciplines, including programming, level design, asset creation and testing.
Hy4 Goes Beyond Simple Code Generation
Traditional coding assistants generally help developers write individual functions, debug errors or generate code from specifications.
Hy4 is designed for longer-running development tasks.
Tencent says the model has improved capabilities in understanding, planning, debugging and validation for long-context software engineering tasks. It has also been optimized to improve the visual quality and interaction experience of front-end development.
Conventional AI Coding Vs Hy4
| Traditional Coding Assistant | Hy4 Preview |
|---|---|
| Generates code snippets | Can handle longer development workflows |
| Function-level assistance | Project-level planning |
| Developer runs most tests | Stronger validation capabilities |
| Limited context | Up to 1M-token context |
| Primarily software code | Code, games, office work and research |
| Usually text-focused | Improved front-end interaction and visual quality |
This does not mean Hy4 can independently create every type of production-ready application without human supervision. It is a preview model, and Tencent itself lists limitations.
Tencent Demonstrates Game Engine Integration
Tencent says Hy4 works fluently with game engines.
In demonstrations reported around the launch, the model was connected through MCP to Unreal Engine 5 and used conversational instructions to build a shooting-game demo, including elements such as environment setup, gameplay and level construction.
The demonstration is important because it shows how AI agents could interact with development tools rather than simply producing text outside the development environment.
AI Game Development
| Development Stage | Hy4 Capability |
|---|---|
| Concept | Understand natural-language request |
| Planning | Break complex task into steps |
| Programming | Generate and modify code |
| Environment | Work with game engines |
| Gameplay | Implement mechanics |
| Testing | Inspect results and validate |
| Refinement | Continue through multiple turns |
This workflow is closer to an AI development agent than a conventional chatbot.
Hy4 Can Create A Playable Prototype
Tencent’s official model description specifically says Hy4 can turn a single prompt into a playable prototype.
Developers can then continue refining complex projects through multiple turns.
That distinction matters.
A prototype does not necessarily mean a commercially finished game. Instead, the model can create a functioning starting point that developers can subsequently modify and expand.
From Idea To Prototype
Idea
│
▼
"Build a 3D survival game"
│
▼
Hy4 plans project
│
▼
Code + assets + mechanics
│
▼
Engine integration
│
▼
Testing
│
▼
Playable prototype
│
▼
Developer feedback
│
▼
Further AI-assisted refinement
The ability to iterate through this process could substantially reduce the time required for early-stage experimentation.
Hy4 Is A 770B-Parameter Mixture-Of-Experts Model
Under the hood, Hy4 uses a Mixture-of-Experts (MoE) architecture.
The model has 770 billion total parameters, but only 49 billion are activated for each token.
Tencent’s technical documentation says the architecture contains 78 layers, with 256 routed experts and one shared expert. Eight routed experts are activated for each token alongside the shared expert.
Hy4 Technical Specifications
| Specification | Hy4 Preview |
|---|---|
| Total parameters | 770 billion |
| Active parameters | 49 billion |
| Architecture | Mixture-of-Experts |
| Layers | 78 |
| Routed experts | 256 |
| Shared experts | 1 |
| Routed experts activated/token | 8 |
| Context length | 1 million tokens |
| License | Apache 2.0 |
The sparse architecture allows Hy4 to have a very large overall parameter count without activating the entire model for every token.
One-Million-Token Context Supports Large Projects
Another major feature is Hy4’s 1-million-token context window.
A large context window is particularly useful for software engineering because an AI agent may need to process extensive project information, including source files, documentation, configuration files, logs and previous interactions.
Large Context In Coding
Large codebase
│
├── Source files
├── Documentation
├── Dependencies
├── Configuration
├── Logs
└── Previous instructions
│
▼
1M-token context
│
▼
Better project-level
understanding
The larger context does not eliminate the need for effective project management, but it gives the model more information to work with during long-running tasks.
Hy4 Also Targets App And Front-End Development
Although the game-generation demonstration has attracted significant attention, Tencent is positioning Hy4 as a general productivity model rather than a game-only system.
The model is designed for software engineering and has specifically been optimized for front-end development.
Tencent says Hy4 provides improvements in the visual quality and interaction experience of front-end work.
This could allow users to describe an application interface and have the model generate the underlying implementation while iterating on the design.
Potential App-Building Workflow
| Step | AI-Assisted Task |
|---|---|
| Idea | Understand app requirements |
| Interface | Generate front-end |
| Logic | Build application functions |
| Data | Connect information sources |
| Testing | Identify errors |
| Refinement | Modify based on feedback |
| Delivery | Produce working prototype |
The technology could therefore be useful for rapid prototyping even outside gaming.
Hy4 Can Generate Business Documents And Spreadsheets
Tencent has also trained Hy4 on office and analytical tasks.
The company says the model can process information spread across multiple files and turn it into documents, spreadsheets and presentations.
It is also designed to handle data analysis, equations and financial models.
Hy4 Productivity Capabilities
Multiple files
│
▼
Hy4 analyzes information
│
├── Data analysis
├── Financial models
├── Documents
├── Spreadsheets
└── Presentations
│
▼
Shareable output
This expands the model’s potential use beyond developers and into finance, research and general knowledge work.
Tencent Built Hy4 With Domain Experts
Tencent says it worked with internal experts in software engineering, gaming, finance and security to develop high-quality training data.
The company also co-designed Hy4 with products such as CodeBuddy and WorkBuddy to optimize the model around real-world workflows.
This approach reflects a broader trend in AI development toward training models around actual workplace tasks rather than relying solely on general-purpose benchmarks.
Areas Used For Hy4 Training
| Domain | Focus |
|---|---|
| Software engineering | Coding and development |
| Gaming | Game creation |
| Finance | Analysis and financial models |
| Security | Security-related workflows |
| Office productivity | Documents and spreadsheets |
| Scientific research | Complex reasoning |
The combination is intended to make Hy4 useful for practical work rather than just conversational tasks.
Tencent’s Internal Evaluation Shows Competitive Performance
Tencent conducted a blind evaluation involving 163 internal experts and 203 engineering tasks.
Hy4 preview received an average score of 2.99 out of 4.00, slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94.
Tencent Blind Engineering Evaluation
| Model | Average Score |
|---|---|
| Hy4 Preview | 2.99 / 4.00 |
| Kimi K3 | 2.94 / 4.00 |
| GLM-5.3 | 2.92 / 4.00 |
The results should be interpreted carefully because the test was conducted internally by Tencent and covered its own selected engineering tasks.
Still, the results indicate that Tencent considers Hy4 competitive with other leading open models.
Hy4 Is Released Under Apache 2.0
Tencent has released Hy4 preview under the Apache License 2.0.
The model weights are available through platforms including Hugging Face, where the model repository identifies the license as Apache-2.0.
Apache 2.0 is a permissive open-source license that generally allows developers to use, modify and redistribute licensed software subject to its terms.
Hy4 Availability
| Platform / Service | Availability |
|---|---|
| Hugging Face | Open-weight model |
| Tencent Cloud TokenHub | API access |
| OpenRouter | API access |
| WorkBuddy | Available |
| CodeBuddy | Available |
| Yuanbao | Available |
| ima | Available |
| License | Apache 2.0 |
Tencent said the model would be available free on WorkBuddy and CodeBuddy for two weeks following launch.
Hy4 API Pricing Targets Cost Efficiency
Tencent has also positioned Hy4 as a cost-efficient model.
The company’s listed API pricing is $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million tokens for cache hits.
Hy4 API Pricing
| Usage | Price |
|---|---|
| Input | $0.834 / 1M tokens |
| Output | $2.501 / 1M tokens |
| Cache hit | $0.042 / 1M tokens |
For developers building AI-powered coding or software-development workflows, pricing can become an important factor because agentic applications may generate large numbers of model calls.
Hy4 Is Also Designed For Scientific Research
Game development is only one part of Tencent’s positioning.
The company says Hy4 has improved capabilities in scientific research, including AI research and development, molecular dynamics, condensed-matter physics and fundamental mathematics.
This reflects Tencent’s attempt to build a model that can handle different forms of technical reasoning.
Scientific Research Applications
| Field | Potential Use |
|---|---|
| AI research | Experimentation and analysis |
| Molecular dynamics | Scientific simulation workflows |
| Condensed-matter physics | Research reasoning |
| Mathematics | Complex problem solving |
| Engineering | Technical analysis |
The company says the model has improved at understanding, reasoning through and solving complex research problems.
Hy4 Helped Optimize Its Own Development
One of the more unusual claims from Tencent concerns the model’s role in its own development.
Tencent says Hy4 participated in automated optimization of training methods, data strategies, evaluation frameworks and low-level operators.
The model proposed approaches, ran experiments and iterated based on results, with code, logs and feedback feeding into later rounds.
Tencent describes this as an early-stage recursive self-improvement loop.
Hy4 Development Loop
Model proposes optimization
│
▼
Experiment executed
│
▼
Results collected
│
▼
Code + logs + feedback
│
▼
Next optimization round
│
└──────────────►
The claim should not be interpreted as the model independently developing itself without human or system oversight. It refers to automated participation in selected optimization processes.
Hy4 Optimized Its Inference System
Tencent also says Hy4 autonomously analyzed bottlenecks in its inference system.
It performed multiple rounds of optimization involving areas such as operator fusion and communication optimization.
According to Tencent, these changes increased end-to-end throughput by 31.8% compared with the baseline, with gains maintained across different context lengths and concurrency levels.
Reported Inference Improvement
| Metric | Reported Result |
|---|---|
| Baseline | 100% |
| Hy4 optimized throughput | 131.8% |
| Improvement | 31.8% |
These figures are Tencent’s reported results and have not been independently verified in the cited sources.
Hy4 Still Has Known Limitations
Despite the impressive demonstrations, Tencent describes Hy4 preview as an early version.
The company’s model documentation lists several known issues.
Hy4 can spend longer than necessary reasoning through complex tasks and has a tendency to over-verify its own work.
These behaviors can increase token consumption and latency.
Known Hy4 Limitations
| Limitation | Potential Effect |
|---|---|
| Excessive reasoning | Higher latency |
| Over-verification | Additional token usage |
| Early preview status | Capability still evolving |
| Complex projects | Human supervision remains important |
| Large model size | Significant hardware requirements for self-hosting |
Tencent says there is still substantial room for improvement in both pre-training and post-training.
Building A Full Game Still Requires Human Oversight
The ability to create a playable prototype from a prompt does not mean AI has eliminated game-development teams.
Complex commercial games require art direction, gameplay balancing, quality assurance, performance optimization, intellectual-property management and extensive testing.
AI can accelerate some of these activities, but developers remain responsible for evaluating the output and deciding what enters a production project.
AI-Assisted Game Development
Human idea
│
▼
AI prototype
│
▼
AI-assisted iteration
│
▼
Human testing
│
▼
Optimization
│
▼
Quality assurance
│
▼
Production release
The strongest near-term use case may therefore be faster prototyping rather than fully autonomous game production.
The Bigger Picture
Tencent’s Hy4 preview represents another step toward AI systems that can execute multi-stage development tasks rather than merely generate individual pieces of code. Its ability to turn a natural-language request into a playable game prototype, work with game engines and refine projects through multiple interactions gives developers a new way to move from an idea to a working demonstration.
The model’s broader specifications are equally significant. With 770 billion total parameters, 49 billion active parameters and a 1-million-token context window, Hy4 is positioned as a large open-weight model for coding, office productivity, gaming and scientific research. Its Apache 2.0 release also gives developers access to the model’s weights, while Tencent’s reported $0.834-per-million-token input pricing provides an additional route for users who prefer API access.
Looking Ahead
The most important development to watch will be whether Hy4’s game-generation capabilities translate from impressive prototypes into reliable workflows for larger and more complex projects. The model’s long-context capabilities, game-engine integration and multi-turn refinement could make AI increasingly useful as a development partner, particularly during prototyping, debugging and repetitive implementation work. However, Tencent’s own documentation makes clear that Hy4 is still an early preview with known reasoning and verification inefficiencies.
For the wider AI industry, Hy4 also reinforces the growing competition around open-weight models and agentic coding. Tencent is combining a very large MoE architecture, permissive Apache 2.0 licensing, long-context processing and real-world productivity capabilities in a single system. If future Hy4 releases improve reliability while preserving the ability to execute complex workflows, the technology could make prompt-driven software and game creation a much more practical part of everyday development.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
