Matrix-Game 3.0 - Real-time interactive world models

Skywork AI has launched Matrix-Game 3.0, a significant open-source release that pushes the boundaries of real-time AI-generated worlds. Available under a permissive MIT license on Hugging Face, the core 5-billion parameter model can generate interactive, consistent environments at 720p resolution and 40 frames per second. A key technical achievement is its 'minute-long memory consistency,' allowing the AI to maintain coherent world states and object interactions over extended sequences, a challenge for previous generative models.

The model was trained on a hybrid dataset combining synthetic data from Unreal Engine and high-quality assets from AAA video games, alongside real-world visual data. This diverse training corpus helps it generate more realistic and physically plausible environments. Architecturally, Matrix-Game 3.0 is designed to scale, with a Mixture of Experts (MoE) framework that allows it to expand up to 28 billion parameters. This structure enables more complex reasoning and detail without a linear increase in computational cost during inference.

For developers and researchers, this release democratizes access to high-fidelity world simulation technology. It provides a foundational tool for rapidly prototyping game levels, simulating training environments for robotics, or creating dynamic backdrops for interactive media. The real-time performance at 40 FPS makes it suitable for applications requiring immediate feedback, moving AI world-building from a slow, offline rendering process to an interactive experience.