Open real-time interactive world model
Matrix-Game 3.0
A streaming keyboard-and-mouse world model designed for real-time interaction and long-horizon memory.
01 / Overview
Streaming interactive video, keyboard and mouse conditioning, long-horizon memory, local open-model research.
A streaming keyboard-and-mouse world model designed for real-time interaction and long-horizon memory.
Record both responsiveness and spatial drift over longer navigation sequences.
Do not compare reported speed across projects without matching hardware and resolution.
02 / Strengths
Where it stands out
- Open implementation and model artifacts make the release inspectable.
- Targets both real-time response and longer-range world consistency.
- Supports direct action control rather than fixed prompt-to-video generation.
03 / Boundaries
What not to overclaim
- Installation and GPU requirements make it unsuitable for most casual visitors.
- Long-horizon memory does not guarantee persistent game objects, rules, or geometry.
- Published performance depends on the documented test configuration.
An open generative interaction stack, not a supported consumer web game or 3D asset creator.
04 / Access & output
How the model reaches people
Official GitHub repository, project page, technical report, and model artifacts for local experimentation.
- Platforms
- GitHub · Local Python environment · Model checkpoints
- Outputs
- Streaming generated video world
- Inputs & compatibility
- Keyboard and mouse action control · NVIDIA CUDA environment
05 / Practical path
How to evaluate it
- 01
Start with official project examples to understand the supported interaction.
- 02
Match the repository instructions and checkpoints specifically to version 3.0.
- 03
Confirm GPU memory and dependency requirements before downloading assets.
- 04
Test revisiting and repeated controls to assess memory claims.
Primary evidence