Google DeepMind releases world model Genie 4

On 28 October 2026, Google DeepMind released its world model Genie 4 as a research preview. According to TechCrunch, the model can generate more realistic interactive worlds, supports real-time scene changes via prompts entered during simulation, and can generate matching real-time audio, including music, ambient sound and interactive dialogue; the report says scene consistency lasts up to two minutes, twice that of its predecessor Genie 3. On the same day, Ars Technica reported that Genie 4 outperforms Genie 3 in almost every respect and uses an updated Gemini model to improve simulation, generating explorable virtual worlds frame by frame in real time from instructions without pre-rendered environments or coded interactive elements; DeepMind sees it as a key step towards AGI, usable for building training environments for physical robots, and said paying users may be able to use Genie 4 soon. The Decoder reported the same day that Genie 4 computes scenes frame by frame and responds to user input, with users able to move via keyboard and change events with text prompts. The Verge reported the same day that the official list gives four main updates: more realistic worlds, real-time modification of the experience with natural-language prompts, audio generation, and use of Gemini 3.8 Flash under the hood to assist simulation. Developments so far centre on the release itself, with no follow-up reports yet.

Written by AI from the reports · updated 58 min ago

Not enough continuous observations yet to draw a trend.