Revealed in August 2025, Genie 3 takes a basic text or image prompt and instantly generates a playable 3D world that is complete with objects that you can move, weather that shifts with commands, and environments that remember what you’ve done, even when you walk away. We're talking 720p visuals, 24 FPS performance, and persistent memory over several minutes of continuous, glitch-free exploration.
Unlike Genie 2, which was impressive but limited to short, grainy video loops, Genie 3 is built for immersion. It supports real-time editing on the fly, just type in “spawn a storm” or “build a cave,” and it happens instantly, no reload required. This level of interactivity is powered by what DeepMind calls an “autoregressive world model,” which isn’t hardcoded with rules. Instead, Genie 3 learns how the world works, gravity, water, and shadows just by watching video data. That means the system doesn’t fake physics; it internalizes them, leading to emergent, realistic behaviour without manual programming.
What really elevates Genie 3 is its spatiotemporal consistency. If you paint a wall or drop a sword somewhere, leave the scene, and return, the AI remembers the state exactly as you left it. That’s a massive step toward AI that understands continuity, something even big game engines struggle with. DeepMind isn’t pitching this as a toy; they see Genie 3 as a training ground for general-purpose intelligence. These hyper-realistic, memory-rich environments are where future AI agents can learn safely, without risking real-world consequences.
Despite its potential, Genie 3 isn’t open to the public yet. It’s currently in limited research preview, accessible only to a select group of developers and researchers while DeepMind fine-tunes its safety and governance protocols.
Still, the implications are crystal clear.
Genie 3 is no longer just about creative play; it’s a foundational step toward artificial general intelligence (AGI), offering a simulated world where machines can learn, adapt, and possibly outpace human intuition. Simply put, Genie 3 doesn’t just build worlds; it builds the infrastructure for AI to truly live in them.
Unlike Genie 2, which was impressive but limited to short, grainy video loops, Genie 3 is built for immersion. It supports real-time editing on the fly, just type in “spawn a storm” or “build a cave,” and it happens instantly, no reload required. This level of interactivity is powered by what DeepMind calls an “autoregressive world model,” which isn’t hardcoded with rules. Instead, Genie 3 learns how the world works, gravity, water, and shadows just by watching video data. That means the system doesn’t fake physics; it internalizes them, leading to emergent, realistic behaviour without manual programming.
What really elevates Genie 3 is its spatiotemporal consistency. If you paint a wall or drop a sword somewhere, leave the scene, and return, the AI remembers the state exactly as you left it. That’s a massive step toward AI that understands continuity, something even big game engines struggle with. DeepMind isn’t pitching this as a toy; they see Genie 3 as a training ground for general-purpose intelligence. These hyper-realistic, memory-rich environments are where future AI agents can learn safely, without risking real-world consequences.
Despite its potential, Genie 3 isn’t open to the public yet. It’s currently in limited research preview, accessible only to a select group of developers and researchers while DeepMind fine-tunes its safety and governance protocols.
Still, the implications are crystal clear.
Genie 3 is no longer just about creative play; it’s a foundational step toward artificial general intelligence (AGI), offering a simulated world where machines can learn, adapt, and possibly outpace human intuition. Simply put, Genie 3 doesn’t just build worlds; it builds the infrastructure for AI to truly live in them.
