Genie 3: A new frontier for world models

TL;DR

Google DeepMind’s August 2025 research-preview announcement of Genie 3 — a general-purpose, text-prompted, real-time interactive world model that auto-regressively generates dynamic 720p / 24 fps environments and maintains consistency over several minutes with visual memory extending back roughly one minute. Genie 3 introduces promptable world events (text-conditioned mutations of the running world, e.g. weather changes or new objects/characters), and is positioned as a stepping stone toward AGI by serving as an unlimited curriculum for embodied agents — the post explicitly shows the SIMA agent goal-pursuing inside Genie-3-generated worlds. Like the later Project Genie productization, this post is qualitative: no architecture, no parameter counts, no benchmarks, no quantitative comparisons.

Key claims

Genie 3 generates dynamic, navigable worlds from a text prompt at 720p / 24 fps in real time, with consistency “for a few minutes” [§intro].
It is DeepMind’s first world model with real-time interaction; positioned alongside Genie 1 (2024 foundation WM) and Genie 2 (large-scale foundation WM) in a single product lineage [§Towards world simulation].
Real-time interactivity is the technical bottleneck: each auto-regressively generated frame must condition on a trajectory that grows with time, and that computation must run multiple times per second in response to live user input [§Pushing the frontier of real-time capabilities].
Visual memory extends back ~1 minute — if the user revisits a location after a minute the model must retrieve relevant information from that far back [§Pushing the frontier of real-time capabilities].
Long-horizon consistency is claimed as an emergent capability of the auto-regressive frame-by-frame generator, in explicit contrast to NeRF / Gaussian-Splatting methods that depend on an explicit 3D representation — the worlds are “far more dynamic and rich because they’re created frame by frame based on the world description and actions by the user” [§Environmental consistency over a long horizon].
Promptable world events are introduced as a second interaction modality on top of navigation — text prompts during a running world can alter weather, introduce new objects/characters, or push counterfactual scenarios [§Promptable world events].
Genie 3 was tested with SIMA (DeepMind’s generalist 3D-environment agent) — the agent is given a goal, sends navigation actions to Genie 3, and Genie 3 simulates the future from the agent’s actions without knowing the goal [§Fueling embodied agent research].
The paper names five concrete capabilities the model demonstrates: modelling physical properties (water, lighting), simulating the natural world (ecosystems, plant/animal behaviors), animation and fiction, exploring real-world locations and historical settings, and the promptable-events mechanism [§Genie 3’s capabilities].
Stated limitations: limited agent action space (promptable events are environment-level, not agent-level), poor multi-agent interaction modeling, inability to faithfully simulate real-world locations with geographic accuracy, text rendering is poor unless the text is in the input world description, and continuous interaction is bounded at “a few minutes” rather than hours [§Limitations].
Released as a limited research preview to a small cohort of academics and creators, with no weights, no architecture, no benchmarks, and no quantitative comparisons disclosed [§Responsibility].

Method

Not disclosed. The post describes only that Genie 3 is an auto-regressive frame-by-frame generator that conditions each new frame on the prior trajectory, supports both navigational inputs and text-based promptable-event inputs, and runs at 24 fps / 720p. There is no mention of architecture (transformer / DiT / something else), no parameter count, no training data scale, no training compute, no benchmarks, and no comparison to Genie 1 / Genie 2 / Veo 3 / external systems beyond the qualitative contrast with NeRF and Gaussian Splatting on the consistency dimension. The acknowledgments list ~30 named research/engineering contributors plus a long second tier of internal collaborators — consistent with a flagship-team-scale effort but uninformative about the model itself.

Results

No quantitative results. The post relies entirely on embedded demo videos across five capability categories (physical properties, natural world, animation/fiction, real locations/historical settings, promptable events), plus a separate SIMA-agent-inside-Genie-3 demo. The only operating-point numbers in the entire post are the four target specs: 720p resolution, 24 fps, “few minutes” consistency, ~1-minute visual memory horizon.

Why it’s interesting

This is the foundational artifact behind the world-model push at DeepMind — the Project Genie product launch in January 2026 (Project Genie: Experimenting with infinite, interactive worlds) is a productization of this model, and Waymo’s Genie-3-derived 4D driving simulator (The Waymo World Model: A New Frontier For Autonomous Driving Simulation) is a post-trained vertical specialization of it. Filing Genie 3 itself anchors that lineage: it is the first announcement of (a) real-time auto-regressive world-model generation at 720p / 24 fps with multi-minute consistency, and (b) promptable world events as a second interaction modality on top of navigation. The “emergent consistency from frame-by-frame AR generation without an explicit 3D representation” framing is also the strongest articulation on file of the World Foundation Models thesis that Sitzmann’s “bitter lesson of CV” essay endorses — Genie 3 is the empirical existence proof Sitzmann’s normative argument leans on. For Luma, this is also the closed-source benchmark target: the open Wan-based long-horizon AR pipeline (Context Forcing: Consistent Autoregressive Video Generation with Long Context, Pretraining Frame Preservation in Autoregressive Video Memory Compression, MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation) cites it implicitly as the bar to clear, and the open-source counterpart Advancing Open-source World Models (LingBot-World) (LingBot-World) names Genie 3 directly.