Matrix-Game 3.0

Matrix-Game 3.0 from Skywork AI (Kunlun Tech): `prompt` and `image` in; `matrix-game-3.mp4` out; runner verified; start on `matrix-game-3`.
Matrix-Game 3.0 is the third Skywork generation, pushing interactive world video to higher resolution with few-step inference. In this catalog the runner is verified — the official cityscape image-conditioned demo runs with actions forward + camera_r, 12 iterations, 3 inference steps, 704x1280, int8 quantization, compiled VAE, and the mg_lightvae decoder.


The paper is arXiv 2604.08995. Project page: matrix-game-v3.github.io. Code: SkyworkAI/Matrix-Game.
Build your run
MatrixGame3Pipelineworldfoundry.pipelines.matrix_game.pipeline_matrix_game_3official-cityscape-imageworldfoundry.pipelineworldfoundry-eval run \
matrix-game-3 \
--pipeline.task-profile official-cityscape-image \
--jsonWhat it is for
Interactive world on Matrix-Game 3. Sibling cards cover 1, 2.0, and 3.5 first-person.
Launch worldfoundry-eval run matrix-game-3 with --pipeline.task-profile official-cityscape-image.
Score on WorldModelBench — action-conditioned game-world rollout quality.
Runtime contract
Required: prompt and image; optional: video; actions: keyboard, mouse; artifacts: matrix-game-3.mp4.
Defaults: num_inference_steps=3, fps=17.
Recorded official-cityscape-image defaults: num_inference_steps=3, fps=17.
Run in WorldFoundry
Status: Runner parity pending. Unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage Skywork/Matrix-Game-3.0 locally — missing files fail at load, nothing is downloaded. Run worldfoundry-eval run matrix-game-3 with --pipeline.task-profile official-cityscape-image. Catalog VRAM is unset.


