Qwen4-Max: Alibaba's new AI model builds 3D games in Unreal and Unity on its own
Alibaba has unveiled Qwen4-Max, a huge AI model with 2.4 trillion parameters that it says can build games in Unreal Engine, Unity and Godot, then test and fix them itself. Alibaba also showed demos from factory planning, financial analysis, law and medicine. Little of it has been independently verified so far, and you can't download the model.

What's going on
On November 16, 2026, Chinese tech giant Alibaba unveiled its new flagship model, Qwen4-Max. It succeeds Qwen3.8-Max and understands images and video as well as text. It can process very long inputs and work on projects that run for several hours. The demo getting the most attention is one where, according to Alibaba, the model turns a plain-language description into a finished 3D game. To do that, it works directly inside game engines like Unreal Engine, Unity and Godot, the same tools professional studios use.
Qwen4-Max is only available through Alibaba's cloud services: the QwenCloud API, the Token Plan, Qwen Studio, QwenWork and the Qoder development environment. Alibaba isn't releasing the weights of the large model. According to ProPakistani, smaller Qwen4 models will be released for download later.
How the AI builds and tests a game
Typical coding assistants write code, and you have to integrate and test it yourself. According to Inc Russia, Qwen4-Max goes further. The model sets up the project structure and writes the code. It places objects in the scene and configures materials, lighting, the user interface and visual effects. For 3D models and scenes it also uses the free 3D software Blender, and according to ProPakistani, it uses it for 3D printing templates too.
The key piece is automatic self-checking. Once a project is done, the agent (the AI that carries out several steps in a row on its own) launches the game. It looks at what's on screen, checks how the game responds to input, and then fixes what's wrong. Alibaba gives these examples:
- a badly placed camera
- broken animations
- poor lighting
- missing feedback when you hit an enemy
According to Kompas, the model can also adjust how the camera follows the player character and change the game's pacing and flow. Because the model tests and corrects itself, it can run several build-and-test cycles without a human constantly stepping in.
The three showcase games
Alibaba showed three projects:
- Meow of Magic (Unreal Engine): You play a spellcasting cat, cast spells, pick upgrades and fight waves of enemies all the way to a final boss.
- Yun Guan Xing (Unreal Engine): A fantasy adventure where the character flies on a sword. It's set in imperial China.
- Flying Racer (Unity 6): A mobile racing game with flying cars, drifting, boost and other mechanics. Alibaba says it could be exported as an installable Android app.
Kompas writes that the games look decent at first glance and the animations run fairly smoothly. Its editors recommend writing instructions that are as precise and complete as possible if you want the result to match what you have in mind.
Live Delivery: work in progress instead of one final result
For long tasks, Alibaba also released Live Delivery, an open-source tool under the permissive Apache 2.0 license. With it, the agent delivers interim versions, gets feedback and improves the result step by step. That means you can look at an early version, change the requirements and have the agent pick up right from that point. At launch, the tool supports Codex CLI, a command-line coding tool, on macOS and Linux.
Tech and numbers
Qwen4-Max is a mixture-of-experts (MoE) model. It's made up of many specialized subnetworks, and only a small share of them work on any given word. Of its 2.4 trillion total parameters, 71 billion are active per token. According to ProPakistani, there are another 328 billion parameters that only pick up patterns between neighboring tokens. For long inputs, the model uses Gated DeltaNet, a leaner alternative to the classic attention mechanism, and sparse attention, where each word only attends to selected other words. Four so-called gated residual branches control how information flows between the network's layers. The context window, meaning the amount of text the model can keep in view at once, is 1 million tokens.
Alibaba says the base model gets by with 25% fewer active parameters than its predecessor. Even so, its "bits per byte" score dropped by 5 to 13% on knowledge, code, question-answering and STEM tests. The lower that number, the better the model predicts text. ProPakistani points out that two of the comparison models had extra training, which can skew the comparison.
| Metric (vendor figures) | Qwen3.8-Max | Qwen4-Max |
|---|---|---|
| Work360 score | 47.74 | 52.69 |
| Avg. cost per task (Work360) | $4.43 | $1.67 |
| Secure functions (functional + secure) | 22% | 31.8% |
| Security review across functions, files, repos | 8% | 14.7% |
Work360 is Alibaba's own test for office and professional tasks. It covers 26 domains, 182 subdomains and 721 specialty areas. At high reasoning effort ("High"), the model used 35% fewer tokens there, and at very high reasoning effort ("Xhigh"), 24% fewer. On the coding benchmark Terminal-Bench 4.0, the figures were 49% and 48% fewer. Inc Russia quotes Alibaba as saying the model uses about 35% fewer tokens on enterprise tasks and costs around 60% less per task.
More demos: factories, finance, medicine
Games are only part of the presentation. According to ProPakistani, Alibaba also showed:
- Factory planning: At the companies SANF and Pacific Union, the model turned 2D plans into 3D production lines for the auto industry. It simulated cycle times, tested control systems, designed parts and checked milling paths. For one part, manual programming time reportedly dropped from 475 to 24 minutes.
- Video: With WonderClip, it planned a video from storyboard through camera work to rendering and editing.
- Finance: Using 90 tools, it gathered market data and analyzed derivatives on China's CSI 300 stock index. The output was an Excel workbook with source links and a linked PowerPoint presentation.
- Law: It sorted 198 documents from IPO preparations and wrote an 86-page due diligence report with supporting evidence.
- Medicine: In an intensive care demo, it analyzed six days of patient records with 1,433 images. Alibaba says its findings matched the treating physician's diagnosis.
- Research: On the Mira platform, it screened 2.7 million molecules in 48 hours for a coolant for high-performance chips. In an Alzheimer's demo, it came up with 24 drug candidates. The selected compound has only been tested in simulation so far and still has to be tested in the lab.
Pricing and access
The API supports web search, tool calling and structured output in fixed formats like JSON. One million input tokens cost $1.50, and one million output tokens cost $4.50. Previously processed inputs that the system caches cost just $0.075 per million. According to ProPakistani, the price of this automatic caching also dropped 70% compared with Qwen3.8-Max.
Our take
Qwen4-Max shows where the big models are headed: away from chatbots and toward agents that operate real software, look at their own output and fix it. For small game studios, that could cut the cost of prototypes. Building core mechanics, testing user interfaces and fixing technical bugs eats up a lot of time today. The price also stands out. It's well below that of many Western frontier models.
But a lot of this is unproven. Almost all the numbers come from Alibaba itself, often from its own tests like Work360. Inc Russia stresses that Alibaba only showed short demo videos, with no source code or playable versions of the games. An independent study called CEGame-Bench reached a sobering conclusion in October: AI agents often manage to build game projects, but they can only play their own games all the way through in 16% of attempts. Qwen4-Max wasn't tested in that study.
And a prototype isn't a finished game. Big commercial titles need balanced game design, solid performance, a distinct artistic style and complex systems, sustained over months or years. Whether the AI can keep up on projects that long will be the real test. If you want to experiment yourself, you'll have to use the cloud or wait for the smaller Qwen4 models Alibaba has announced.
Related videos
Sources
- Google News (RU) incrussia.ru
- propakistani.pk propakistani.pk
- tekno.kompas.com tekno.kompas.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.