What is Matrix-Game 4.0?
Matrix-Game 4.0 is Skywork AI's open-weight model for playable game worlds, released on 15 October 2026 with inference code, a browser demo, game source code and weights. Unlike earlier world models that try to imagine the whole game, it leaves the gameplay to an ordinary game engine: the engine runs the physics, the controls and the rules, draws a plain low-poly picture, and a causal video model repaints that picture into detailed frames a few at a time. Skywork calls this streaming neural rendering.
Each scene has its own checkpoint of about 5 billion parameters, fine-tuned from Wan2.2-Fun-5B-Control, which is itself built on Alibaba's Wan2.2 5B video model. Three scenes ship: Medieval Castle and Ancient Egypt, both third-person shooters, and City Racing, a driving game. Output is 864 by 480 pixels.
The weights, sink images and game assets are Apache-2.0; the Wan and umT5 parts keep their own Apache-2.0 terms and the fast TAEHV codec is MIT. Training code and training data are not released.
Is Matrix-Game 4.0 worth it?
Matrix-Game 4.0 fixes the problem that has made video world models feel like dreams: nothing stays put. Earlier systems, including Skywork's own Matrix-Game 3.0 from March, predicted every frame from the last one, so a wall could melt or a corridor rearrange itself if you turned round. Version 4.0 keeps the world in a real game engine. Collisions, scores and physics are ordinary code, and the 5B model only repaints what the engine draws. A door you walked through is still there because the engine says so.
Speed is what makes it a game rather than a clip. Skywork's GitHub page reports up to 67 frames per second on an RTX 5080 and 58 on an RTX 5070 Ti at 864 by 480, with fixed prompts and the optional SageAttention backend, measured at steady state without video input and output. Those are consumer cards, and optimised paths exist for the A100 and RTX 5090 as well. The model card is more cautious: it promises no frame rate and no fit on consumer memory, and it publishes no VRAM figure, so the GitHub numbers are the vendor's best case.
The catch: it is a renderer for three scenes, not a game generator. Each checkpoint works only with its own low-poly scene and camera profile; you cannot point it at your Unity project. Skywork lists drift in appearance and geometry errors with new geometry, fast camera moves, occlusion and long sessions, and the neural image never feeds back into gameplay, so what you see can briefly disagree with what you hit. Without training code, adapting it to a new game means rebuilding the pipeline yourself.
The licence is the cleanest in the category. Apache-2.0 on weights and assets allows commercial use with no revenue cap or territory carve-out, which is more than Tencent's HY-World 2.0 offers. Adoption is still small: 59 likes and 984 downloads in the last month on Hugging Face, and 2.8k stars on GitHub.
Our take: this is the architecture to watch rather than a tool to ship with. If you build games, the lesson is that neural rendering on top of a normal engine already runs at playable speed on a mid-range card; the day someone releases the training recipe, it becomes a style filter for any game.
Where it earns its place
A graphics engineer at a mid-sized game studio
Judge whether neural rendering could replace hand-made art for a stylised level
Run the Medieval Castle scene on an RTX 5080 and compare the low-poly engine frames against the repainted output for latency and drift. Because the code and assets are Apache-2.0, the prototype can sit in an internal tool without legal review of a custom licence.
A PhD student working on world models
Reproduce a real-time baseline that keeps state consistent
The technical report (arXiv 2610.27999) and the three released checkpoints give a fixed reference: a 5B Wan2.2-based model, two denoising steps in the quality preset and one in the realtime preset, and an attention cache of 21 latent frames.
A creative technologist at an agency
Show a client an interactive, playable demo in a browser
The released browser demo streams the City Racing scene; the client drives, and the engine handles collisions so the demo does not break in ways a pure video model would.
Before you commit
- Setup
- Clone SkyworkAI/Matrix-Game, follow the v4 README in the Matrix-Game-4 folder, and run the manifest downloader with --repo-id Skywork/Matrix-Game-4.0; it checks each file's SHA256 and places the three scenes under resources/models. The weights are not drop-in Diffusers pipelines.
- Integrations
- Bundled game engines for the three scenes, an online browser demo, and optimised inference paths for A100, RTX 5090, RTX 5080 and RTX 5070 Ti (FP8 linear layers, fused kernels, CUDA graphs, optional SageAttention).
- Data and privacy
- Runs entirely on your own GPU. Weights, sink images and game assets Apache-2.0; TAEHV codec MIT; racing audio includes CC0 and CC BY 3.0 material whose credits file must be kept.
- Cost at scale
- No API or hosted service; the cost is one GPU per player session. Skywork's fastest published figure is on an RTX 5080.
- Limits
- Each checkpoint is tied to its own scene and camera profile; 864x480 output; drift with new geometry, fast camera motion, occlusion or long sessions; neural output does not affect collisions; no training code or data.
What’s good
- Game logic stays in a real engine, so the world does not rearrange itself
- Up to 67 fps on an RTX 5080 at 864x480 (vendor figure, steady state)
- Apache-2.0 weights and assets, commercial use allowed
- Inference code, browser demo and game source all released
What’s not
- Works only with its three bundled scenes and camera profiles
- No training code or data, so new games need your own pipeline
- Drift and geometry errors on fast motion, occlusion and long sessions
- No published VRAM requirement; model card promises no frame rate
Best for
Game and graphics researchers who want a working, permissively licensed example of real-time neural rendering on consumer GPUs.
Matrix-Game 4.0 alternatives
Matrix-Game 4.0: common questions
- Is Matrix-Game 4.0 worth it?
- Matrix-Game 4.0 is the first open world model that plays like a game, because it stopped trying to be the game and became the graphics card's painter instead. We rate it 3.9 out of 5.
- How much does Matrix-Game 4.0 cost?
- Free: weights, game assets and inference code under Apache-2.0. You pay for the GPU.
- What are the downsides of Matrix-Game 4.0?
- Works only with its three bundled scenes and camera profiles; no training code or data, so new games need your own pipeline; drift and geometry errors on fast motion, occlusion and long sessions; no published VRAM requirement; model card promises no frame rate.
- Who is Matrix-Game 4.0 best for?
- Game and graphics researchers who want a working, permissively licensed example of real-time neural rendering on consumer GPUs.
- How much does Matrix-Game 4.0 cost for a team or in production?
- No API or hosted service; the cost is one GPU per player session. Skywork's fastest published figure is on an RTX 5080.
- Is Matrix-Game 4.0 safe for company data?
- Runs entirely on your own GPU. Weights, sink images and game assets Apache-2.0; TAEHV codec MIT; racing audio includes CC0 and CC BY 3.0 material whose credits file must be kept.
- What are the best Matrix-Game 4.0 alternatives?
- The closest alternatives we have reviewed are HY-World 2.0, LTX-2.5 and Kling 4.0.
Answers checked Nov 5, 2026.