Real-time playable world model from Skywork pairing a game engine and coding agent with a few-step causal video model.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Matrix-Game 4.0 is an interactive world model from Skywork that wraps a video model in a harness which holds the game state explicitly. A game engine evolves everything the rules can determine and renders it as a low-poly proxy view. A coding agent handles cases the rules do not cover: it turns user instructions into new rules registered in the engine, and reads engine state to write dynamic prompts.
The video model then only renders. It takes the per-frame proxy, a dynamic caption and a few reference frames, and generates one latent (four frames) per step using a causal DiT distilled to one or two denoising steps, with a rolling KV cache. The project page shows continuous play sessions of roughly 54 to 99 seconds across several games, with HUD and key inputs.
Creator coverage reports per-scene models of about 5B parameters at 480x864, around 83 fps on a single consumer GPU, and weights and inference code released under Apache 2.0. These figures come from video reviews and were not confirmed on the project page text. Output is 480p-class video rather than true 3D assets.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Splits work between a game engine that renders a blocky low-poly scene and a video model that repaints it, so world state persists. Per-scene ~5B param models at 480x864, up to about 90 seconds, weights and inference code under Apache 2.0.
World model that creates playable real-time 3D games with consistent mechanics. A game engine handles physics and rules while the AI renders detailed visuals; includes a coding agent and runs at 83 fps 480p on one consumer GPU.
Closest neighbours by meaning, not just by category.
Model
Real-time multiplayer world model that generates four-player Rocket League-style gameplay, playable online.
Model
Real-time interactive world model with compressed memory and hybrid action and semantic control, built on Wan 2.2.
Research
Research framework where a coding agent maintains world state as code, rendered to video by a fine-tuned MiniMax-H3 model.
Model
Open framework and dataset for turning video diffusion models into real-time, long-horizon interactive world models.
Model · Nvidia
NVIDIA research model that generates real-time, action-controlled multi-agent video of shared worlds, from two to four players.
Model
Interactive autoregressive world model generating minute-long playable video with real-time camera control and prompt events.
