World models
A world model is a neural network that generates a scene you can move through and steer as it responds to your inputs, instead of running game code. The best-known is Google DeepMind's Genie 4, announced in an October 27, 2026, blog post. DeepMind says it generates music, background sound and dialogue along with visuals, and that game developers among its testers use it to turn concepts into explorable prototypes before writing engine code. TechCrunch reported it keeps a scene consistent for up to two minutes, twice as long as Genie 3. For now it's a research preview for trusted testers (our story).
Others differ. Reka says its 19-billion-parameter Rho-1 can act as a steerable world model, but its announcement caps video at 672×384 and says long streams lose their layout before they lose visual detail (our story). Viggle's JST-2, released October 19, works in 3D: Viggle says it generates a Gaussian splat scene in about one second, and its Scene API returns .spz files for $0.06 a scene from a prompt (our story).
None produces game logic you can ship, and DeepMind says Genie 4's details and object permanence can drift. They suit prototyping, or helper roles next to an engine, such as AnyGameRenderer, a model from a paper by Ruicong Liu and colleagues that repaints gray-box gameplay video into polished visuals (our story). Our AI game engines guide compares them.
What to watch: DeepMind says Genie 4 is coming soon to Google AI Ultra subscribers in Project Genie.
Sources
Sources for the introduction above.
- Google DeepMind: Genie 4: A new generation of world simulation
- TechCrunch (Rebecca Bellan): Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering
- Reka: Rho-1, collapsing the multimodal stack
- Viggle: JST-2
- Viggle: Developers / Scene API
- Hugging Face Papers (Ruicong Liu et al.): AnyGameRenderer: Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark
Written in our own words; quotes are marked and credited. How we report
