Skywork AI, a unit of Kunlun Tech, open-sourced Matrix-Game 4.0 on October 15. This version takes a different approach: the game engine owns the world state, the physics rules and player input, while the video model only has to redraw the engine's low-poly output as a detailed picture.
The headline figures from the developers are 480p at 60 fps on a single RTX 5090, with latency from player action to on-screen response below 150 milliseconds. The project describes it as streaming neural rendering for playable games.
The engine handles logic, the model handles visuals
According to the description on GitHub, the video model in 4.0 is a causal, few-step generative model. Its input is low-poly RGB frames rendered by the engine plus a text prompt, and its output is a continuous stream of highly detailed frames. Game rules, collision and scoring all stay inside the engine.
There is also a coding agent. Players ask for a change in natural language, such as a different gameplay rule, and the agent translates it into new game rules and rendering prompts, so the visual style and the gameplay can change together.
The release includes:
- Model weights, inference configuration and inference code (single-GPU inference)
- A browser demo and the game source code
- Assets for three scenes: two third-person shooters set in a medieval castle and ancient Egypt, plus a city racing game
The README lists the license as Apache-2.0, while third-party components and assets keep their own licenses. The developers have not disclosed the model's parameter count.
The road across four versions
Lined up side by side, the changes across the Matrix-Game versions are clear.
Version 1.0, released in May 2025, had more than 17 billion parameters and focused on Minecraft-style scenes. The model generated the next stretch of video directly from player actions, and the world itself was also imagined by the model. Version 2.0, in August 2025, put the emphasis on low latency and long interactive sequences, generating continuously at around 25 frames per second. Version 3.0, released on March 27 this year, added long-term memory and targeted high resolution and real-time interaction, positioned against Google's Genie 3.
In all three earlier versions the model carried both the world and the picture, and consistency was a constant problem: walk far away, turn back, and the scene might no longer match. Version 4.0 hands the world back to the engine and leaves the model with rendering only, so consistency is backstopped by the engine, and the frame rate rises from roughly 25 fps in 2.0 to 60 fps. The trade-off is that what can be played returns to what the engine can express. Players cannot conjure an object or rule that does not exist in the engine and can only change it through the coding agent.
The approach is close in spirit to neural rendering such as Nvidia's DLSS: the traditional pipeline produces a low-fidelity result and a network fills in the detail. The difference is that Matrix-Game 4.0 fills in far more, with the low-poly input supplying little beyond composition and motion.
Developers in China who want to reproduce it face a hardware hurdle. Because of export controls, the officially sold card in China is the cut-down RTX 5090D, while the project's published test setup is the RTX 5090. The README gives no data on what frame rate the D variant or the previous-generation 4090 can reach, or whether latency would exceed 150 milliseconds. For small teams hoping to prototype with it, running the official demo on their own card first is more reliable than trusting the headline numbers.
The developers list three directions for future work: higher resolution, support for more game genres, and keeping what each player sees consistent in multiplayer scenes. 480p is still some way from mainstream game visuals, and no third-party benchmarks on other graphics cards are available yet.
Sources: SkyworkAI GitHub project page, TechNode, CocoLoop, Pandaily. Resolution, frame rate and latency follow the single-RTX 5090 test setup published by Kunlun Tech; the license follows the project README.