Matrix-Game 4.0

Skywork AI

Streaming Neural Rendering for Playable Games

Project Website | Technical Report | Models & Game Assets | Code & Quick Start | Previous Versions

Matrix-Game is Skywork AI's series of open-source interactive world models. Matrix-Game 4.0 combines executable game engines with a causal video model: the engine handles gameplay, physics and player input, while the model transforms low-poly RGB frames into detailed visuals through streaming neural rendering.

https://github.com/user-attachments/assets/4ded08da-1309-43c9-9662-6e2f2b3fbb41

News

• October 16, 2026: Optimized inference for RTX 5080 and RTX 5070 Ti is now available. With fixed prompts and SageAttention, model throughput reaches up to 67 FPS on RTX 5080 and 58 FPS on RTX 5070 Ti at 864 x 480 (steady state, excluding video I/O).
• October 16, 2026: The Matrix-Game 4.0 technical report is now available.
• October 15, 2026: RTX 5090 optimized inference is now available, with FP8 linear layers, fused kernels, CUDA graphs and an optional SageAttention backend.
• October 15, 2026: Released Matrix-Game 4.0. Inference, browser demo and game source code are now available, with model weights and game assets for Medieval Castle, Ancient Egypt and City Racing on Hugging Face.
• March 27, 2026: Released Matrix-Game 3.0, a real-time streaming interactive world model with long-horizon memory.
• August 12, 2025: Released Matrix-Game 2.0, an interactive world foundation model for real-time long video generation.
• May 12, 2025: Released Matrix-Game 1.0, the first open-source release of the Matrix-Game series.

Matrix-Game 4.0

• Engine-grounded gameplay: game logic and physical state remain in an executable engine, separate from neural rendering.
• Streaming generation: a causal, few-step video model renders successive chunks conditioned on low-poly RGB and text prompts.
• Three scene models: Medieval Castle and Ancient Egypt for third-person shooting, and City Racing for driving.

Start with the v4 README for installation, resource downloads, offline rendering and the online browser demo. The release includes the three native games and single-GPU inference paths for A100, RTX 5090, RTX 5080 and RTX 5070 Ti.

Game assets, sink images and inference configuration files are included in the Hugging Face repository. See the project website for demos and the broader project overview.

Previous Versions

Earlier implementations remain available in their original directories:

• Matrix-Game 3.0: streaming interaction with long-horizon memory.
• Matrix-Game 2.0: real-time interactive long video generation.
• Matrix-Game 1.0: the first Matrix-Game world model.

Citation

If you use Matrix-Game 4.0 in your research, please cite:

@misc{zhang2026matrixgame40harnessingvideo,
      title={Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds},
      author={Jihai Zhang and Runjia Qian and Kaiqin Kong and Biao Jiang and Puyi Wang and Siru Chen and Tianle Li and Guanjie Chen and Junyi Wu and Chubin Chen and Yuanhang Li and Kaichen Huang and Haobo Zhang and Yangguang Li and Wei Yu and Guibin Chen and Yu Cheng and Yahui Zhou},
      year={2026},
      eprint={2610.27999},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.27999},
}

License

Matrix-Game 4.0 is released under Apache-2.0; see the v4 license. Third-party components and assets retain their respective licenses. Earlier versions retain their existing license terms; see the repository's MIT license and the notices within each version directory.