Embodied AI Glossary中文

Matrix-Game (Skywork)

昆仑万维 Matrix-GameAdvanced

Skywork's open-source real-time interactive world model that generates game footage frame by frame from keyboard and mouse input.

Matrix-Game is a series of interactive world models open-sourced by Skywork AI, under Kunlun Wanwei; version 1.0 was released in May 2025, and 2.0 in August 2025. The model reads the current frame plus the user's keyboard and mouse input at each frame, then generates the next segment of video, producing a game world generated as you play. 2.0's main ideas are three: automatically collecting about 1,200 hours of interaction-annotated video using Unreal Engine and the GTA5 environment; injecting keyboard-and-mouse actions into the model frame by frame; and distilling the diffusion model into a few-step causal autoregressive generator that produces minutes-long video in real time at 25 FPS, with weights and code both open-sourced. Version 3.0, released in late March 2026, added camera-position-based long-term memory, reaching 720p at 40 FPS with a 5B model. This family of models follows the same line as Genie 3, and is seen as a potential tool for robot simulation and data generation.

ExampleIn a GTA-style driving scene, holding W and moving the mouse right makes the model generate, in real time, footage of the car driving forward and turning right.

Also called
Matrix-Game 2.0, Matrix-Game 3.0
Related
Interactive World Model · World Model · Genie 3 · Autoregressive Video Generation · Self Forcing · Diffusion Step Distillation
Sources
arXiv 2508.13009: Matrix-Game 2.0
GitHub: SkyworkAI/Matrix-Game
arXiv 2604.08995: Matrix-Game 3.0
As of
2026-04

See it in the full glossary →