Alibaba showed an AI that itself creates games in Unreal Engine and Unity, launches them and corrects errors

Источник изображения: ChatGPT

Alibaba talked about the capabilities of its recently introduced Qwen4-Max AI model in game development and 3D modeling. The model can work with the game engines Unreal Engine, Unity and Godot, create game scenes, program mechanics and independently check the result, making corrections after the project is launched.

  Image source: ChatGPT

Image source: ChatGPT

The peculiarity of the Qwen4-Max approach is that the model is not limited to writing program code. It runs the game itself, analyzes visual feedback, and fixes any problems it finds, including issues with the camera, character animation, lighting, and game effects.

Three games are given as an example. Meow of Magic (adventure with a magic cat), Yun Guan Xing (fantasy sword flying game) and Flying Racer. The first two were created on Unreal Engine, the third – on Unity 6, which we managed to assemble as an installable application for Android. In addition, Qwen4-Max can work with Blender: the model can create 3D objects, assemble scenes, adjust materials and lighting, and then render.

  Image source: orcarouter.ai

Image source: orcarouter.ai

To organize large projects using AI agents, Alibaba introduced an open source Live Delivery tool under the Apache 2.0 license. The tool allows you to gradually release working versions of the project, check the results of each stage and adjust further work without having to wait for the entire task to be completed. The tool currently supports the Codex CLI console, macOS and Linux.

Qwen4-Max is built on an architecture with 2.4 trillion parameters, of which 71 billion are activated when processing each token. The context window reaches 1 million tokens, which allows the model to work with large projects and long sequences of actions. The usage cost is $1.50 per million input and $4.50 per million output tokens. Compared to the previous Qwen3.8-Max, tariffs have decreased from $2 and $6, respectively.

According to the independent platform Artificial Analysis, Qwen4-Max scored 42.89 points in the Intelligence Index versus 33.65 for the previous Qwen3.8-Max. In the Terminal-Bench 4.0 test, which evaluates AI’s ability to perform multi-step tasks using tools, the result was 48.5% versus 38.9%. In FrontierCode-AA 2.0, the model received 26.5% versus 23.4%, and in Terminal-Bench-Science – 23.3% versus 11.9%. On the HLE-Diamond test the result reached 44.9%. However, these indicators do not directly measure the ability to create games, but only indirectly characterize the skills of programming and performing complex sequences of actions.

  Image source: orcarouter.ai

Image source: orcarouter.ai

At the same time, the presented game projects still exist mainly in the form of demonstration videos, and Alibaba has not published their source code and finished assemblies for independent verification. Therefore, Qwen4-Max’s ability to create full-fledged commercial games is not yet clear.

If you notice an error, select it with the mouse and press CTRL+ENTER.