Editorial illustration for DeepMind Genie 4 Modernizes Pre-LLM Agents Using Gemini Omni
DeepMind Genie 4: AI Agents Now Build Worlds in Real Time
DeepMind Genie 4 Modernizes Pre-LLM Agents Using Gemini Omni
Google DeepMind rolled out Genie 4 this week, the follow-up to Genie 3, which the company unveiled back in August 2025. Unlike a video generator that renders a clip and hands it over, Genie 4 builds its world frame by frame, reacting as a user moves through it. Press the keyboard, type a prompt, and the scene adjusts on the fly.
The real shift is what's running underneath. Gemini models now sit in the background, planning multistep tasks and deciding what happens next in the simulation rather than just responding to the last input. That's a return to a design pattern from older agent research, where a planning layer directs a reactive system, dressed up with the kind of real-time generation that wasn't possible before large language models existed.
DeepMind is also pointing to a first: Genie 4 generates matching audio on its own, wind through trees, footsteps, characters talking to each other and reacting as the scene unfolds. The company is upfront that longer sessions and complicated physics still trip it up. For now, access is limited to a small group of testers, with a wider rollout planned for Google AI Ultra subscribers later.
Once they mature, world models could serve as training environments where robots and AI agents learn from simulated experience rather than just from data. Deepmind sees pairing Gemini with Genie as an early step toward hybrid world models that combine the reasoning of language and code models with pixel-level simulation.
Why this matters
Genie 4 is Deepmind pairing an old idea, the agent that plans inside a world model, with Gemini's video and reasoning stack. That combination matters more than the demo footage of weather changing on command. For years, world models and LLM agents have run on separate tracks: one simulates physics and visuals, the other reasons about goals and language. Deepmind is now using Gemini Omni to generate the frames and audio while other Gemini models handle the planning layer underneath, which suggests the company sees simulation and reasoning as two halves of the same product rather than competing research bets.
For developers and founders, the real signal is architectural: multistep planning and real-time environment generation are being built to work together, not bolted on after the fact. That's a different bet than most "world model" demos we've seen, which tend to show off visuals and skip the planning problem entirely. Deepmind calling this "experimental" is the right caveat. Coherence over short clips is one thing; holding a simulated world together across long, user-directed sessions is the actual test, and we haven't seen that yet.
Related Reading
Browse all Google news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks
From Archives CoRover Lists Three BharatGPT Models on Government-Backed AIKosh Platform
From Archives Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models
From Archives Google's Genie 4 world model adds audio, realistic worlds, real-time steering
Open Source
From Archives Anthropic Plans USD 269 Billion in Spending With Microsoft, Google, Amazon
Business & Startups
Common Questions Answered
How does Genie 4 differ from traditional video generators in its approach to world simulation?
Unlike traditional video generators that render a complete clip upfront, Genie 4 builds its world frame by frame while reacting to user inputs in real-time. When users press keyboard controls or type prompts, the scene adjusts dynamically on the fly, creating an interactive simulation experience rather than a static pre-rendered video.
What role do Gemini models play in Genie 4's architecture?
Gemini models run in the background of Genie 4, handling critical planning functions and deciding what happens next in the simulation. Rather than relying solely on video generation, Gemini's reasoning capabilities enable multistep task planning and intelligent decision-making that drives the interactive world model.
How could Genie 4's world models benefit robot and AI agent training according to DeepMind?
Once mature, Genie 4's world models could serve as training environments where robots and AI agents learn from simulated experience rather than relying exclusively on real-world data. This approach allows agents to develop skills and behaviors in controlled, virtual environments before deployment in physical settings.
What is significant about DeepMind combining world models with LLM agents in Genie 4?
For years, world models and LLM agents operated on separate tracks—one simulating physics and visuals while the other reasoned about goals and language. Genie 4 represents a breakthrough by integrating these two approaches through Gemini Omni, which generates frames and audio while other Gemini models handle the planning layer, creating a unified hybrid system.
Further Reading
- Genie 4: A new generation of world simulation - Google DeepMind
- We’re announcing Genie 4, our new world model for real-time interactive simulation - Google Blog
- Genie 3: A new frontier for world models - Google DeepMind
- Genie 4's real upgrade isn't the graphics. It's the second model watching them. - DEV Community
- Project Genie: Experimenting with infinite, interactive worlds - Google Blog
