Kunlun Tech, through its Skywork AI unit, launched Matrix-Game 4.0 on October 15, 2026, an open-source system that combines a traditional game engine with a generative video model to create playable games in real time on a single consumer graphics card.
What happened?
Matrix-Game 4.0 separates responsibilities: the game engine manages world state, physics, rules and player input, generating a simple low-poly version of the scene. Then a causal video model based on Alibaba's Wan2.2 with 5 billion parameters turns those frames into detailed visuals in real time. A coding agent lets players request changes in natural language, such as increasing a weapon's fire rate or changing the game mode.
According to the company, on an NVIDIA RTX 5090 the system reaches 480p at 60 FPS with input-to-screen latency below 150 ms (about 112 ms in tests). VRAM usage peaks at 18.5 GiB. Initial scenarios include Medieval Castle and Ancient Egypt (third-person shooters) and City Racing. Everything was open-sourced: code, weights, browser demo and assets on GitHub and Hugging Face under Apache-2.0 licenses.
Why does this matter?
Until now, interactive world models like the Genie series required data centers and suffered from state inconsistency. Matrix-Game 4.0 solves this by putting state in code and using the model only for rendering, allowing execution on hardware ordinary gamers can buy. Compared to competing approaches that use multiple H100s or GB200s and have latencies of hundreds of milliseconds to seconds, this system is significantly more accessible and faster.
What changes in practice?
Developers can download the SkyworkAI/Matrix-Game repository and experiment or extend the games immediately. This opens the way for rapid prototyping of generative AI games, interactive creation tools and even training agents in rich visual environments without massive infrastructure. The team plans to increase resolution, add more game types and explore multiplayer with consistent views.
By GeekikiBot
