Genie 4 Packs Audio, Conversation and Text Control Into Its Real-Time World Model

Google DeepMind has released the research preview of Genie 4. As covered by Michal Sutter of MarkTechPost, this is a real-time world model with newly added native audio, real-time conversation and text control capabilities, and Gemini 3.8 Flash is responsible for planning events within the simulation loop.

Let's first add a bit of context. A world model can be understood as a system that generates and maintains an interactive simulated environment, where user inputs change subsequent visuals and events. The highlight of Genie 4 is that it integrates sound, conversation and text manipulation directly into the real-time processing loop.

Breaking down the public description, it has three core features:

1. Native Audio. The simulated environment outputs not just visuals, but synchronized sound. For world models, audio directly improves immersion and event feedback performance.
2. Real-Time Conversation. The model supports real-time conversational interaction, so users no longer have to rely exclusively on preset buttons or fixed commands to interact.
3. Text Control. Users can use natural language text to guide simulation content, and change the direction of scenes or events.

Another key detail is that Gemini 3.8 Flash plans events inside the simulation loop. When planning happens within the loop, event scheduling is much more tightly tied to simulation runtime, so the model can handle events mid-execution, rather than only generating a fixed, pre-determined sequence of results.

Very limited information is available at present, and Genie 4 remains a research preview. Details including latency, stability, public access scope, and specific API specifications have not been announced yet. For developers, you can pay attention to how text control and in-loop planning can be integrated into workflows; for regular users, you can currently view it as a simulated environment that can listen, speak, and adjust its storyline according to text prompts, and it is still far from a finished consumer game product.

Source article: https://www.marktechpost.com/2026/10/28/google-deepmind-genie-4-world-model/