AI News Feed
Market watch
Infinigence AI Leads NVIDIA GPU Benchmark as Chinese Firms Push Proactive AISamsung May Add a Second Wide Foldable Alongside Galaxy Z Fold 9Tianjin University’s Zhang Changqing presents trusted multimodal fusion research at CNCC 2026Doubao Work Adds Recordable Computer Skills and Lock-Screen Task ExecutionSK Hynix Earnings Lift Asian Tech Shares as AI Trade ReturnsPolymathic AI and Princeton Researchers Release FlashMap for Faster LLM Decoding
Infinigence AI Leads NVIDIA GPU Benchmark as Chinese Firms Push Proactive AISamsung May Add a Second Wide Foldable Alongside Galaxy Z Fold 9Tianjin University’s Zhang Changqing presents trusted multimodal fusion research at CNCC 2026Doubao Work Adds Recordable Computer Skills and Lock-Screen Task ExecutionSK Hynix Earnings Lift Asian Tech Shares as AI Trade ReturnsPolymathic AI and Princeton Researchers Release FlashMap for Faster LLM Decoding
Large Language Models

Google DeepMind Releases Genie 4, a Real-Time World Model With Native Audio and a Gemini Planner

Google DeepMind has released Genie 4, a real-time world model that builds explorable, audio-enabled scenes from text and image prompts and runs a Gemini model inside the simulation loop as a planner.

AI News Feed · 2026-10-28 18:16

Genie 4 produces ambient noise, music and character dialogue alongside its visuals, making it DeepMind's first real-time world model to generate audio natively, according to the company. Users move through scenes with keyboard controls and steer events with plain-language prompts, changing the weather, moving the camera, or adding a person or object mid-session. DeepMind says scenes simulate physics and human motion, citing aerodynamics and collisions, and calls the behaviour of characters reacting to their surroundings the biggest leap over the previous version.

In the runtime loop described by DeepMind, Gemini interprets a typed prompt and adapts it to the running scene, Gemini 3.8 Flash breaks a goal into a sequence of steps, incoming frames are reviewed to confirm each step succeeded, and the plan is revised if the world shifts or the user intervenes. Between user inputs, Gemini also steers the scene so the world keeps progressing. A separate Gemini Omni model anchors the look and behaviour of each world.

In one research demo, a Gemini model split a game of checkers into steps, each of which was passed to Genie 4 as a text caption; Genie 4 then simulated Apptronik's Apollo humanoid playing the game. DeepMind presents such agent and robot training environments as a main use case.

Genie 4 succeeds Genie 3, announced in August 2025, which ran at 720p and 24 frames per second with roughly one minute of visual memory. DeepMind did not disclose resolution or frame rate for Genie 4 and published no benchmark tables. TechCrunch, briefed on the model, reported scene consistency of up to two minutes, double that of Genie 3; DeepMind's own materials say sessions last a few minutes and that objects drift after repeated prompt edits.

Among other disclosed specifications, Runway's GWM Worlds 2, released on 3 September 2026, runs at 720p and 24 frames per second with native 48 kHz audio including speech, while Skywork's Matrix-Game 3.0, released on 27 March 2026, has 5 billion parameters plus a 2x14B mixture-of-experts variant, reaches up to 40 frames per second and is distributed as open weights under an MIT licence.

Genie 4 is available to trusted testers now and is expected to reach Google AI Ultra subscribers in Project Genie soon, followed by early access partners of the Gemini API and Google AI Studio. All generated video and audio carry SynthID watermarks, and DeepMind says it ran internal safety evaluations before the preview. Testers so far span creative industries, journalism, education, gaming and engineering; reported uses include filmmakers running multiple live takes for previsualisation, photojournalists turning a single archival photograph into a simulation, game developers testing styles and mechanics before writing engine code, and enterprise partners generating zero-to-one simulation scenarios.

DeepMind lists several open problems on the model page: visual details and object permanence drift in longer sessions, after complex prompt edits, or during navigation, and interaction is limited to a few minutes rather than extended sessions.