Tasks and inputs
Matrix-Game 4.0 is listed for Text-to-video, Video-to-video. It accepts video, image and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
Matrix-Game 4.0 is Skywork's open set of 5B real-time video renderers that turn a running game's frames into clean styled video, per scene.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No API pricing stated; open weights for self-hosting
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Earlier world models tried to be the whole game: physics, input handling and pictures in one network. Matrix-Game 4.0 gives up on that. A real, executable game decides what happens. The model only re-renders what the game produces into clean low-poly RGB video, frame-aligned with a separate HUD and sound layer.
Skywork released it on 15 October 2026 with three scene-specialised checkpoints, each with 5,001,016,512 BF16 parameters: Medieval Castle, Ancient Egypt and City Racing. It builds on Wan2.2-Fun-5B-Control. Inference code, a browser demo and the game source came out together.
Because gameplay is deterministic code, the old failure modes of neural game engines (objects that vanish, rules that drift) move out of the model's hands. What remains is a rendering problem with a strict contract:
Castle and Egypt use an independent identity sink; Racing uses the supplied image as frame zero. Skywork is explicit that these policies are not interchangeable.
| Item | Detail |
|---|---|
| Hub repo | Skywork/Matrix-Game-4.0 |
| Pipeline | Matrix-Game 4.0 inference source only; not a Diffusers drop-in |
| Shared parts | T5 encoder, tokenizer and Wan2.2 VAE pulled from the base repo |
| Integrity | Manifest downloader checks length and SHA256 of every file |
| Not included | Training code and training data |
| Licence | Apache-2.0 |
On 16 October Skywork added tuned inference for RTX 5080 and RTX 5070 Ti cards. The Hub repository predates the public release by three days.
Matrix-Game 3.0, from March 2026, was a streaming interactive world model with long-horizon memory. Version 4.0 narrows scope to rendering per scene in exchange for game-grade consistency. Choose 3.0 for open-ended world exploration and 4.0 when you need controllable, repeatable gameplay.
Studios prototyping a neural art pass over grey-box levels should try it: feed your own capture and judge temporal stability over a long session. For new scenes you would need your own training, which this release does not cover. More from the lab is under Skywork's model listing, and related tools sit in the video-to-video directory.
Compare the model on the work you need to complete.
Matrix-Game 4.0 is listed for Text-to-video, Video-to-video. It accepts video, image and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
Cosmos3-Nano-Sim-Auto is NVIDIA's open causal driving simulator: real-time multi-camera video and LiDAR that react to an AV policy's planned path.
NVIDIA
Cosmos3-Nano-Sim-Transfer is NVIDIA's open 16B model that renders simulator depth, segmentation or edge maps into photorealistic robot video in real time.
NVIDIA
Cosmos3-Nano-Transfer-Auto is NVIDIA's open 16B model that generates synchronised multi-camera driving video and LiDAR from map controls and a prompt.
Tencent
WeVisCut is Tencent's 9B tool-calling agent that edits narration scripts into multi-minute videos through 12 editing tools and an editable Studio timeline.
skywork
SkyReels-V3-V2V-14B is Skywork's 14B video-extension model that continues a clip for 5-30 seconds or cuts to a new shot, at 720p from a text instruction.
AInVFX
AInVFX/SeedVR2_comfyUI is a source-linked Hugging Face repository assigned to Video-to-video. Its recorded task interface accepts video and produces video.