Game studios spend years building open worlds by hand. A research team in Beijing wants the world to generate itself while you walk through it — and to hand the recipe to everyone else for free.
The announcement
On July 19, 2026, during the World AI Conference (世界人工智能大会) in Shanghai, Kunlun Tech (昆仑万维) — the Chinese listed company behind the Skywork model family — held a forum on world models and multimodal paradigms. Chairman and CEO Fang Han used the stage to declare 2026 the "year of the world model" (世界模型元年), and Skywork chief scientist Cheng Yu presented the company's new Matrix-Game 3.5.
The model was released with open weights, continuing a strategy the company began a year earlier: publish the interactive world model, not just the demo video.
What Matrix-Game 3.5 can do
A world model is different from a video model. Instead of generating a fixed clip, it generates an environment you can steer — camera, character, events — frame by frame, in real time. The company's stated specifications for version 3.5:
- A 5-billion-parameter model that runs at 720p resolution.
- Real-time interactive generation at 20 frames per second on a single GPU.
- A one-minute memory window, so the world stays consistent as you move through it.
- Patch-level spatial memory: past frames are stored as 3D-coordinate patches and re-injected into the diffusion transformer, replacing cruder frame-level memory.
The company also disclosed its training data scale — more than five million high-quality video clips, over ten thousand effective training hours, and coverage of more than 1,200 game scenes — figures that come from its own pipeline and cannot yet be independently verified.
The lineage matters
Matrix-Game 3.5 did not appear from nowhere. On August 12, 2025, Kunlun open-sourced Matrix-Game 2.0, a 1.8-billion-parameter model that generated minute-level interactive video at 25 FPS and was described in Chinese financial media as the first open-source solution achieving real-time long-sequence interactive generation in general scenes. The 3.5 release is the second step of a deliberate cadence: prove the idea, open it, then scale memory and fidelity.
What steering a world actually feels like
The difference between a video model and a world model is easiest to feel in the controls. In a generated clip, the camera path is fixed the moment rendering starts; in Matrix-Game, the user keeps issuing commands while the world answers, and the next frame depends on what just happened. Turn left and the terrain that appears must be consistent with the terrain you already saw — which is exactly what the one-minute memory window protects.
That requirement explains the engineering choices. Frame-level memory blurs as scenes grow; patch-level memory, which stores history as re-projectable 3D fragments, keeps distant views recognizable when the camera returns to them. For players, that translates into the small but decisive sense that the world persists when you are not looking at it — the property that separates an explorable space from a looping screensaver.
Why games are the first battleground
Fang Han's argument at the forum was blunt: open-world games will be the first industry changed by world models, because content that once required hundreds of people working for years can be generated on demand. Within three to five years, he predicted, world models will become infrastructure for the game industry.
The same machinery serves adjacent markets:
- Embodied AI research, where world models act as simulation engines for training robots before they touch reality.
- Film and virtual production, where directors can scout generated environments instead of building sets.
- Education and simulation, where explorable worlds are cheaper to generate than to model.
According to Li An, chief scientist at BrainNet (脑机网), open-sourcing is the strategic move here: when real-time world generation becomes commoditized, value shifts from who owns the model to who owns the content pipeline and the player community built on top of it.
Key takeaways
- Matrix-Game 3.5 is a 5B-parameter open-source world model: 720p, 20 FPS real-time, one-minute memory, single-GPU.
- It replaces frame-level memory with patch-level 3D spatial memory for better consistency.
- Kunlun's open-source cadence began with Matrix-Game 2.0 in August 2025 and continues with 3.5.
- Games, embodied AI and virtual production are the intended first markets.
Honest limitations
The headline numbers are company-reported; 20 FPS at 720p with one-minute memory remains far from what a modern game engine delivers, and interaction is limited to simple controls rather than full game logic. "Open-source" releases of this kind often omit training code and data pipelines, so reproducibility is partial. And the claim that world models will rebuild game development within five years is a projection from an interested party, not an industry consensus.
What readers can do now
- If you are technically inclined, download the open weights and run the model locally; a single consumer GPU is enough to feel the difference between watching a video and steering a world.
- Indie developers should prototype one small explorable scene and evaluate whether the visual consistency holds across a real gameplay loop.
- Watch the memory specification, not the FPS number — how long a world model remembers is what determines whether it can hold a game, not just a clip.