Genie 3

Genie 3 is a general-purpose foundation world model developed by Google DeepMind and announced on August 5, 2025, that generates interactive, navigable 3D environments from a single text prompt and runs in real time at 720p resolution and 24 frames per second while maintaining environmental consistency for several minutes.^[1][2] Google DeepMind's announcement calls Genie 3 "a new frontier for world models" and describes world models as "a key stepping stone on the path to AGI," positioning it primarily as a training substrate for general-purpose AI agents.^[1] It is the direct successor to Genie 2 (December 2024), improving on it in resolution, latency, real-time interactivity, and the duration of consistent memory.^[1][15]

Unlike passive video generators such as Sora or Veo 3, Genie 3 responds to user navigation inputs, allowing real-time exploration and modification of the worlds it creates.^[1] Users can steer a character through generated landscapes, trigger weather changes, introduce new objects, and revisit previously seen locations, which the model recalls from a roughly one-minute memory window.^[1] DeepMind says the system does not rely on a hard-coded physics engine and instead teaches itself how objects move, fall and interact.^[4]

Genie 3 became publicly accessible through Project Genie, an experimental research prototype, on January 29, 2026, initially limited to Google AI Ultra subscribers in the United States, before expanding to AI Ultra subscribers worldwide in May 2026.^[6][33] On October 27, 2026 Google DeepMind announced a successor, Genie 4, released first as a research preview for trusted testers.^[27]

What is Genie 3?

At a technical level, Genie 3 sits at the intersection of three previously distinct research areas. The first is video generation, where models like Sora, Veo 3, and Runway's Gen models have produced increasingly fluent short clips from text. The second is reinforcement learning with simulated environments, where game engines like Unity or Unreal have long provided diverse training grounds for agents. The third is neural rendering, including NeRF and Gaussian splatting, which produces navigable 3D scenes from images. Genie 3 borrows ideas from all three but does not fit cleanly into any of them: it generates video in real time, supports interactive navigation like a game engine, and produces persistent scenes like a 3D reconstruction system, while operating on neither stored geometry nor a deterministic physics simulator.^[5]

The model is best understood as a learned simulator. It does not encode worlds as meshes, textures, or geometry. Instead, it has internalized statistical patterns about how scenes look, how they change when a viewer moves, and how objects behave over time. Each frame is the model's best guess at what should come next, given everything that has happened in the session so far.^[5]

This approach has two large consequences. It scales with model capacity rather than scene complexity, which means richer worlds do not require richer engineering. It is also fundamentally probabilistic, which means the same prompt and the same actions can produce slightly different sessions on different runs. For a creative tool, that variability is a feature; for an agent training environment that demands repeatability, it is a known constraint that practitioners work around.

How did the Genie family evolve?

Genie 1 (February 2024)

The Genie series began with a research paper published by Google DeepMind in February 2024, titled "Genie: Generative Interactive Environments," authored by Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, and colleagues.^[3] The original Genie model was a foundation world model trained on roughly 30,000 hours of internet gameplay footage from hundreds of 2D platformer games, using a curated dataset of approximately 6.8 million 16-second video clips at 10 frames per second and 160x90 pixel resolution.^[3]

Genie 1's architecture had three primary components: a spatiotemporal video tokenizer, a latent action model, and an autoregressive dynamics model.^[3] All three components used spatiotemporal transformers as their core building block. The latent action model employed a VQ-VAE approach, learning a discrete codebook of latent actions (such as MOVE_RIGHT) without any labeled action data, a technique called unsupervised action learning.^[3] The dynamics model then predicted the next video frame conditioned on the current frame and the inferred latent action.^[3]

Genie 1 scaled from 40 million to 2.7 billion parameters during development, with the final published model reaching 11 billion parameters.^[3] However, it generated worlds at approximately one frame per second, far too slow for interactive play, and operated only in two dimensions.^[3] Its primary value was demonstrating that a model could learn controllable world dynamics from unlabeled video footage alone.^[3]

The Genie 1 paper was notable for proposing that controllable world models could emerge from passive video data, a claim that had been treated skeptically before. By recovering an action codebook from pixel transitions, the model showed that the act of moving a character produces distinctive visual signatures that a network can identify and then replicate. Researchers cited this as evidence that internet video is a richer training substrate than previously assumed, an argument that has since been picked up by other labs.

Genie 2 (December 2024)

Google DeepMind released Genie 2 in December 2024 as a substantial upgrade, extending world generation to three dimensions and improving fidelity.^[15] Genie 2 could accept a single image or short text prompt as input and generate a navigable 3D environment from it.^[15]

However, Genie 2 had notable constraints. Its effective memory window was approximately 10 to 20 seconds, meaning the world would become inconsistent or visually degrade beyond that point. DeepMind said Genie 2 could keep worlds consistent for up to a minute, but most of the examples it published lasted 10 to 20 seconds.^[34] Resolution was capped at 360p, well below broadcast quality. The system was not available for public use and remained a research demonstration.

Genie 2 was significant nonetheless because it showed that the latent action approach from Genie 1 could generalize to 3D environments and photorealistic imagery, not just stylized 2D game footage. Internally, DeepMind treated Genie 2 as evidence that scaling the approach was worth pursuing, which set the trajectory for Genie 3 to focus on real-time throughput and consistency rather than fundamental architecture changes.

Genie 2 also introduced the idea of "world generation as content creation" within DeepMind's communication. Where Genie 1 was framed as a stepping stone for agent learning, Genie 2 was presented with applications including game prototyping, creative ideation, and rapid environmental authoring in mind. That framing would expand further with Genie 3.

Genie 3 (August 2025)

Genie 3 was announced on August 5, 2025 through a Google DeepMind blog post, an accompanying model page, and a limited research preview program.^[1][2] Demis Hassabis, CEO of Google DeepMind, promoted the release, which DeepMind's announcement titled "a new frontier for world models",^[1] and positioned it within DeepMind's broader thesis that learned simulators would become a foundational substrate for agentic AI.

The announcement focused on three headline capabilities: real-time generation at 720p and 24 FPS, environmental consistency over several minutes, and promptable mid-session events.^[1] Together these turned the system from a video curiosity into something closer to a working interactive medium. DeepMind released hand-picked demo footage showing users skiing through snowy landscapes, walking through pastoral scenes, and exploring fictional architectures, all generated entirely from text.^[1]

Genie 3 was not made publicly available in August 2025. DeepMind ran a limited research preview, giving early access to a small cohort of academics and creators.^[1] Project Genie, the consumer-facing launch, did not arrive until January 29, 2026.^[6]

Successor: Genie 4 (October 2026)

Google DeepMind announced Genie 4 on October 27, 2026 and released it as a research preview for trusted testers, saying it was coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio.^[27] DeepMind described it as its first real-time world model to generate audio natively, including interactive dialogue.^[27] Thurrott reported that Genie 4 would raise Project Genie's per-world exploration limit from the 60 seconds available under Genie 3 to 2 minutes.^[28]

What can Genie 3 do?

Real-time interactive generation

Genie 3's headline advance over its predecessors is that it runs at interactive speeds. The model generates environments at 24 frames per second and 720p resolution, which is sufficient for fluid gameplay.^[1] Users navigate generated worlds with standard keyboard controls, and the system responds within fractions of a second to each input.^[1]

This is technically non-trivial. Autoregressive models generate each output token (in this case, each video frame) by conditioning on everything generated previously. As a session continues, the model must reference an ever-growing trajectory of prior frames and actions. DeepMind's announcement post, written by Jack Parker-Holder and Shlomi Fruchter, described the challenge: "The model has to take into account the previously generated trajectory that grows with time."^[1] Achieving real-time throughput under this growing computational burden required substantial engineering work alongside the base model architecture.

The result is a system where each frame is generated on-the-fly based on the player's current action and the accumulated history of the session, rather than being retrieved from a pre-rendered sequence or stored game state. The contrast with conventional game engines is sharp. A traditional engine has the scene fully laid out in memory and draws frames from it; Genie 3 has only the model's weights plus the rolling context buffer, and has to reconstruct what the world should look like every 41 milliseconds.

Memory persistence

Genie 3 maintains a visual memory window of approximately one minute.^[1] When a user leaves a location and later returns to it, the model reconstructs it in a way that is consistent with what was generated earlier.^[1] Objects that were moved or created remain in their new positions.^[1] Environmental changes the user triggered persist for as long as the memory window covers the relevant frames.

This one-minute window is several times longer than the 10 to 20 seconds typical of Genie 2's published examples,^[34] and it enables a qualitatively different experience: players can walk away from a location, explore nearby areas, and return to find their starting point roughly intact. Prior world models would simply regenerate everything from scratch upon re-entry, producing visible inconsistencies.

The consistency mechanism does not rely on an explicit 3D scene representation. There is no NeRF, no Gaussian splatting, and no stored point cloud. Instead, consistency emerges from the model attending to earlier frames during autoregressive generation.^[5] This means the approach scales with model capacity rather than with scene complexity, but it also means that the "memory" degrades gracefully as the window fills rather than maintaining perfect accuracy indefinitely.

In practice the memory exhibits a few characteristic failure modes. Objects far from the camera tend to drift or change subtle details when revisited. Patterns on textured surfaces, such as the exact arrangement of leaves on a tree or stones on a path, can change between visits while the larger shapes stay constant. Researchers testing the system have noted that the model's "memory" is closer to the gist-level recall that humans report for scenes than to the pixel-perfect persistence of a game engine. For many applications that is acceptable; for some it is not.

Promptable world events

Beyond navigation controls, Genie 3 supports mid-session text prompts that alter the generated world. Users can type instructions like "make it snow," "introduce a thunderstorm," "add a fox near the treeline," or "change it to night," and the model modifies the ongoing generation accordingly. Google DeepMind calls this feature promptable world events, noting that they make it possible to "change the generated world, such as altering weather conditions or introducing new objects and characters," which "increases the range of scenarios agents can use to learn about handling unexpected situations."^[2]

This capability separates Genie 3 from purely action-conditioned video models, which can only respond to movement inputs.^[5] Text conditioning during generation allows users to author their worlds dynamically rather than just explore a fixed prompt output. Researchers studying agent behavior can use this to introduce novel stimuli mid-session without restarting a session from scratch.

The feature has some limitations. Very specific or geometrically precise instructions ("place a red cube exactly three meters to the left of the blue sphere") do not work reliably.^[5] The model interprets prompts in a statistical sense, producing plausible interpretations rather than literal executions.^[5] Subtle effects, like adjusting lighting temperature or producing a specific musical sound, are also unreliable because the model's training data does not cleanly separate cause from effect in those dimensions.

Oasis is fully action-conditioned and provides no parallel text channel during play; Marble is scene-level and does not generate frame-by-frame during interaction at all. Genie 3 occupies the middle ground where both modalities are active throughout.

Self-learned physics

Genie 3 was not given a physics engine during training. DeepMind says it does not rely on a hard-coded physics engine and instead learns how objects move, fall and interact.^[4] Google DeepMind researchers described this as the model discovering consistency mechanics without explicit programming, and it reflects a broader principle that large-scale video pretraining encodes substantial physical knowledge.^[1]

In practice the physics simulation is approximate. TechCrunch noted that in one demo of a skier going down a mountain, the snow did not move as it would in relation to the skier.^[4] Falling objects and basic rigid-body behavior are handled reasonably, but complex fluid dynamics, cloth simulation, and multi-body contact are not reliably correct.^[5] The system is best understood as a plausible physics approximator rather than a ground-truth simulator.^[5]

This distinction matters for use cases. For agent training where the goal is general navigation policies, an approximate physics simulation works because the agent learns to act sensibly in plausible environments rather than to exploit any specific physical bug. For robotics sim-to-real transfer, however, the gap between learned physics and ground truth becomes a significant obstacle, because a robot that learns to walk in Genie 3 may not transfer that policy faithfully to a real robot whose contact dynamics differ.^[5]

Researchers at DeepMind and elsewhere have noted that this is a fundamental property of pixel-level world models, not a Genie 3-specific limitation.^[5] Models like Sora, Pika, and Runway Gen-3 all produce videos where physics is locally plausible but globally inconsistent. Genie 3's interactive framing puts more pressure on consistency, which is why its physics failures stand out even though the underlying model has roughly comparable training characteristics.

Diversity of generated worlds

DeepMind's demonstrations showed Genie 3 generating environments across a wide stylistic and content range. Examples included natural landscapes (forests, mountains, oceans, deserts), urban scenes (city streets, plazas, interiors of buildings), stylized worlds (cartoon vistas, abstract designs, surrealist environments), and historical or fictional settings (medieval villages, fantasy landscapes, science fiction environments).^[1] The model can switch between styles fluently based on text prompting, including art directions such as a watercolor-painting aesthetic.^[2]

This breadth comes from training on diverse video sources, including stock footage, animations, gameplay recordings, and consumer-uploaded content. The model has not memorized specific environments but has learned the visual grammar of how environments of various types should look and behave. Asked for a generic "forest in autumn," it produces a plausible one; asked for "a forest in autumn with red maple leaves" it adjusts; asked for "a forest in autumn in northern Hokkaido," it produces something forest-like with vaguely Japanese affect but cannot recover the actual geography.

How does Genie 3 work?

Genie 3 is an autoregressive transformer model. Like large language models that generate text one token at a time, Genie 3 generates video one frame at a time, with each frame conditioned on all prior frames and the most recent user action.^[2]

The spatiotemporal transformer architecture inherited from Genie 1 remained central to Genie 3, though Google DeepMind has not published a full technical paper describing the specific modifications made between versions.^[2] The model processes both spatial information (what the world looks like in the current frame) and temporal information (how the world has changed over the session) within its attention mechanism, allowing it to maintain coherence across the time dimension.

Researchers familiar with the family have inferred from public material that Genie 3 likely combines a few major architectural ingredients:

• A high-capacity diffusion transformer (DiT) used as the per-frame generator, similar in spirit to the architecture behind Sora and Veo 3, but conditioned on an action sequence rather than only a text prompt.
• A spatiotemporal tokenizer that compresses video frames into latent representations that the autoregressive component can operate on efficiently.
• An attention mechanism that lets the current frame's generation attend to earlier frames within the memory window, supporting both short-range temporal coherence (frame-to-frame smoothness) and long-range consistency (recall after revisiting a location).
• Specialized scheduling and caching machinery to make the autoregressive loop run within the 41 millisecond budget required for 24 FPS, including techniques borrowed from speculative decoding and KV cache reuse in language modeling.

None of these inferences have been confirmed in detail by DeepMind. The blog post says real-time interactivity "required significant technical breakthroughs," but does not enumerate them.^[1]

Training involved large volumes of diverse video footage, extending well beyond the 2D platformer games used for Genie 1. Genie 3 was trained on video spanning many environment types, including natural landscapes, architectural scenes, stylized animation, and fictional settings. This breadth is what allows it to generate "any world," rather than being specialized to a particular visual domain.

At its August 2025 announcement, Genie 3 took only text prompts;^[5] by November 2025 DeepMind described it as generating worlds "from a single image or text prompt,"^[10] and Project Genie accepts uploaded images.^[6] This contrasts with Genie 2, which could bootstrap a world from a single input image.^[15]

Google DeepMind has not disclosed the exact parameter count for Genie 3. Published reports referencing an 11 billion parameter count are drawing on the Genie 1 scaling experiments, which found the best results at 11B.^[3] Whether Genie 3 uses the same scale, or has been scaled further, has not been officially confirmed.

The lack of a formal technical paper at launch was noteworthy. Google DeepMind published a full arXiv preprint for Genie 1, which allowed external researchers to reproduce and analyze the architecture in detail.^[3] Genie 3's announcement consisted of a model page, a blog post, and a limited research preview program.^[1][2] Community discussion on Hacker News and in AI research forums noted that without a paper, key design decisions, training data composition, evaluation benchmarks, and failure modes cannot be independently assessed.^[11] Google DeepMind has not stated whether a technical report will be published.

The inference pipeline runs on Google's internal accelerator infrastructure. Google has not disclosed the hardware configuration used to serve Genie 3 at real-time speeds, though the compute requirements for generating 720p video at 24 FPS autoregressively are substantial. Project Genie's server-side deployment means users access the model through a browser-based interface rather than running it locally, which is consistent with the hardware demands involved.^[6]

The economics of running such a system at consumer scale are unclear. Project Genie was offered only through Google AI Ultra, a plan priced at $249.99 per month when Google introduced it in May 2025.^[6][29]

Training data and supervision

One of the more interesting aspects of Genie 3, inherited from Genie 1, is that it learns from unlabeled video. Most generative video models are trained with text captions paired to clips, which provides a strong signal for what is happening in the frame. Genie 3 adds a layer underneath that: the model not only learns what scenes look like, but also which transitions between frames correspond to plausible "actions" by an implicit camera or character.^[3]

This matters because labeled action data is scarce. Hand-annotating which frames in a video correspond to "walking forward" versus "turning right" is expensive at the scale needed for foundation model training. Genie 3 sidesteps that bottleneck by inferring action structure from the raw frame-to-frame pixel changes, and then learning to generate similar transitions when given navigation inputs at inference time.^[3]

The trade-off is that the learned action space is not perfectly aligned with user-friendly controls. The model knows what "forward motion" looks like, but the precise mapping between a key press and a particular trajectory has to be learned by the user as well, through trial and error during a session. Players often describe Genie 3 navigation as feeling slightly "floaty" or "dreamlike" compared to a traditional first-person game, a property that likely stems from this learned-action approach.

Is Genie 3 publicly available?

On January 29, 2026, Google DeepMind launched Project Genie, an experimental research prototype that puts Genie 3 in the hands of end users.^[6] Project Genie rolled out to Google AI Ultra subscribers in the United States aged 18 and older.^[6] Google AI Ultra was priced at $249.99 per month when Google introduced the plan in May 2025.^[29] On May 19, 2026, Google expanded Project Genie to AI Ultra subscribers worldwide and added an integration with Google Street View (initially for places in the US) as a source for world initialization.^[33]

Project Genie offers three main modes of interaction:

World sketching: Users describe a world with text prompts, optionally augmented with reference images. An integrated tool called Nano Banana Pro provides a preview of the world before the user enters it, allowing adjustments to the initial prompt.^[6]

World exploration: Once inside a generated world, users navigate in real time using WASD keys for movement, arrow keys for camera control, and spacebar for vertical movement.^[7] Genie 3 generates the path ahead continuously as the user moves through it.

World remixing: Users can explore a curated gallery of worlds created by other users, remix them with new prompts, and download video recordings of their sessions.^[6]

At launch, Project Genie capped individual sessions at 60 seconds of generation, shorter than what Genie 3 is capable of in research settings.^[7] TechCrunch reported that the limit was partly due to budget and compute constraints, since each session needs dedicated compute, and DeepMind's Shlomi Fruchter said longer sessions would add little to the value of the testing.^[7] Some capabilities present in the August 2025 Genie 3 announcement, such as full promptable world events, were not included in the initial Project Genie release.

The launch was limited to the US, with Google stating that access would expand "to more territories in due course" without specifying a timeline; that expansion arrived with the worldwide rollout in May 2026.^[6][26]

Reaction to Project Genie's public availability was mixed in tone but enthusiastic in interest. TechCrunch's Rebecca Bellan published a hands-on piece titled "I built marshmallow castles in Google's new AI-world generator," finding the prototype "inconsistent, sometimes impressively generating playable worlds, other times producing baffling results that miss the mark," and the keyboard navigation frustrating.^[7] 9to5Google's Abner Li reported that sessions were limited to 60 seconds, with 720p worlds and interaction rates of 20 to 24 frames per second.^[8]

What is Genie 3 used for?

Agent training

Google DeepMind's stated primary motivation for building Genie 3 is to create a substrate for training embodied AI agents.^[1] The current bottleneck for embodied agent research is the scarcity of diverse training environments. Real-world data collection is slow and expensive; hand-crafted game environments are diverse but narrow in scope; existing 3D simulation platforms like Unity or Unreal require significant engineering work to set up.

Genie 3 can generate an essentially unlimited variety of environments from text prompts, potentially allowing agents to train across far more situations than any manually authored simulator provides.^[1] Researchers can specify the type of environment, the physical properties they want the agent to encounter, and then generate hundreds of variations without writing any code.

Google DeepMind demonstrated this use case through integration with SIMA (Scalable Instructable Multiworld Agent), their embodied AI agent.^[1] SIMA 2, announced in November 2025, was tested in newly generated Genie 3 worlds and was also able to improve itself through self-directed training inside Genie environments.^[10] When placed in newly generated Genie worlds, SIMA 2 could orient itself, interpret user instructions, and take meaningful actions toward goals despite the unfamiliar visual context.^[10]

The goal is eventually to close the sim-to-real gap: train agents in generated worlds diverse enough that their learned behaviors transfer to physical robots operating in the real world. As of early 2026, this transfer remains an open research problem. Genie 3 environments are photorealistic by consumer standards but not reliably accurate enough for robust real-world deployment.

A particular challenge is that current reinforcement learning approaches benefit from massive parallelism, where thousands of agent instances train simultaneously across independent simulator processes. Genie 3's per-instance compute cost makes this kind of parallelism expensive at present. Researchers have discussed two ways to address this: distilling Genie 3 into smaller, faster student models that can be replicated cheaply, and using Genie 3 sessions as offline data sources for policies trained on logged trajectories rather than live interaction. Both approaches are areas of active research.

Autonomous vehicle simulation

In February 2026, Waymo announced the Waymo World Model, a specialized derivative of Genie 3 adapted for autonomous driving research.^[9] The Waymo World Model generates multi-sensor outputs including camera video and lidar point clouds, transferring Genie 3's knowledge from 2D video into 3D lidar outputs.^[9]

The system can simulate rare edge cases that Waymo's actual vehicle fleet would almost never encounter in normal operations: tornadoes, unusual animal encounters, extreme weather, and other low-frequency but high-importance scenarios.^[9] Waymo noted that Genie 3's pretraining on a vast and diverse video corpus gives the Waymo World Model access to visual and physical patterns that Waymo's own data collection would never adequately cover.^[9]

By feeding Waymo's specialized post-training on top of Genie 3's general world knowledge, the company can generate synthetic training data for situations that are nearly impossible to capture safely at scale in the real world.^[9] Bloomberg reported that Waymo believes these simulations will help accelerate the expansion of its robotaxi operations.^[17]

The collaboration is significant because it is a domain-specific application of Genie 3 by a sister Alphabet company rather than by Google DeepMind itself. It also frames the model as a base for vertical specialization, with Waymo essentially treating Genie 3 the way model providers treat language model base checkpoints: a starting point that is fine-tuned with proprietary data for a particular domain. Other domains that have been discussed publicly as candidates for similar specialization include surgical training, logistics planning, and architectural visualization.

Game development prototyping

Game developers have been among the most interested outside observers of Genie 3. The ability to generate a navigable environment from a text description in seconds compresses what would otherwise be days or weeks of asset creation and level design into a nearly instant prototype.

GDC's 2026 State of the Game Industry survey found that 52% of game industry professionals believed generative AI was having a negative impact on the industry (up from 30% a year earlier and 18% the year before),^[36] reflecting tension between AI as a productivity tool and AI as a threat to creative roles. Genie 3 itself attracted a range of reactions. Some game developers saw it as a useful ideation tool, pointing to its ability to quickly rough out visual concepts and spatial layouts. Others noted that a navigable environment generated by Genie 3 is not a game: it has no mechanics, no scoring, no narrative, no optimization for frame delivery, and no tools for iterating on a design.

Julian Togelius, a prominent researcher in AI and games, who tried the research preview, described Genie 3 as a "neural game engine" and by far the most advanced such model, and saw ideation as its viable use, but did not expect it to replace game engines because it could not offer "the very precise and predictable editing you have in a regular game engine."^[14]

Google positions Genie 3 for rapid prototyping and concept exploration, not as a replacement for commercial game development pipelines. As of early 2026, no shipping commercial game has been announced as using Genie 3 directly, though several studios have publicly speculated about hybrid pipelines where Genie 3 generates reference content that is then implemented in traditional engines.

Education and creative exploration

DeepMind cited education and creative exploration as additional use cases.^[1] Students studying history, geography, or science could in principle navigate generated representations of historical locations, ecosystems, or physical phenomena. The imperfect accuracy of Genie 3's world models makes formal educational applications limited today, but the potential for illustrative and exploratory contexts is real.

Creatives including filmmakers and concept artists have used early access to Project Genie to rough out environment ideas, generate reference material for settings, and produce short video clips of generative worlds for use in mood reels and pitch materials.

Robotics research

Beyond Google DeepMind's own SIMA integration, robotics researchers across academia have begun exploring Genie 3 as a possible source of pre-training data for visual policies. The argument runs as follows: training a visual policy on real robot data is expensive and slow, but training on synthetic data generated by traditional graphics tools requires either hand-authored environments or expensive 3D capture. A learned video world model bypasses both bottlenecks, generating training scenes that approximate real-world visual statistics.

The challenge, as in autonomous vehicle simulation, is that the model's physics are not aligned with reality at the level required for a policy to transfer cleanly to a physical robot.^[5] Researchers exploring this approach typically use Genie 3 for visual representation pre-training (learning useful visual features) and then fine-tune the resulting model on a smaller corpus of real robot data, rather than expecting full policy transfer from Genie alone.

A few teams have also looked at using Genie 3 as an environment in which to train high-level planning policies, where the agent reasons about which sequence of subgoals to pursue and a lower-level controller handles execution. In that setup the precise physics matter less; the agent just needs the visual world to behave roughly correctly while it plans its next move.

How does Genie 3 differ from other world models?

By August 2025, Genie 3 entered a competitive field of world model efforts from both large labs and startups.^[16] The approaches differ significantly in technical architecture, output format, and intended use case.

Decart Oasis

Decart AI released Oasis in late October 2024, in collaboration with hardware startup Etched.^[38] Oasis is a playable Minecraft-like world model that runs in real time in the browser.^[20] It runs at approximately 20 frames per second and demonstrated that real-time interactive world generation was achievable before Genie 3's announcement.^[20] Oasis operates at lower resolution and with less diversity than Genie 3, and its output is visually noisier, but its free public availability made it an influential demo.

Oasis was notable for two reasons besides being first. It was optimized for Etched's upcoming transformer-specific AI accelerator chips, although the public demo ran on Nvidia H100 GPUs.^[38] And it focused specifically on Minecraft-like environments, which let it sidestep the general-purpose visual diversity challenge that Genie 3 had to solve. The two approaches are essentially complementary: Oasis showed how to do narrow real-time interactive generation cheaply; Genie 3 showed how to do broad real-time interactive generation at higher fidelity but at greater compute cost.

When Genie 3 launched, Oasis was often cited in commentary as the natural baseline for what "real-time playable world models" looked like in production. The comparison was generally favorable to Genie 3 on fidelity and breadth, and favorable to Oasis on accessibility and compute economics.

Microsoft Muse

Microsoft Research published Muse on February 19, 2025, in a Nature paper titled "World and human action models towards gameplay ideation."^[18][19] Muse was developed in partnership with Xbox Game Studios and trained on extensive gameplay footage from Bleeding Edge, a multiplayer game published by Ninja Theory.^[18] Where Genie 3 generates open-ended worlds from text, Muse predicts game state and pixel-level transitions within a specific game environment for which it has detailed training data.^[18]

The two systems thus solve different problems. Muse is narrow but accurate, modeling a specific game's mechanics and visual style faithfully. Genie 3 is broad but approximate, modeling general worlds plausibly. Muse's research thesis was that game-specific world models could support "gameplay ideation," letting designers explore variations of an existing game without manually coding them;^[18] Genie 3's research thesis is more ambitious, treating the world model as a substrate for any embodied training task.

Microsoft also distinguished its approach by focusing more on the agent side: the Muse paper devoted significant space to a "world and human action model" architecture that jointly models both world transitions and the actions a player would take.^[19] Genie 3 by contrast leaves action choice to a separate agent or human user. This makes Muse arguably more useful for game design exploration and Genie 3 more useful for agent training across diverse environments.

World Labs Marble

Fei-Fei Li's World Labs took a fundamentally different bet with its commercial product Marble, announced in November 2025.^[13] Where Genie 3 generates a continuous video stream that the user navigates, Marble produces downloadable, editable 3D assets and environments from multi-modal inputs including photos, videos, panoramas, and 3D layouts.^[13] An analyst summarized the distinction this way: Marble renders what the world looks like; Genie 3 shows how the world changes. Marble's output is persistent and author-controlled, which makes it more useful for content pipelines but less useful for agent training.

The contrast reflects a deeper architectural choice. World Labs has emphasized that a true "spatial intelligence" system needs an explicit 3D representation, not just a sequence of pixels.^[13] Genie 3 takes the opposite stance: pixel-level prediction is enough if the model is powerful enough. Both companies have framed their respective approaches as the natural path toward general spatial AI, and both have argued that the other approach has fundamental limitations. The debate is unsettled.

In practice the two products serve different markets. Marble is being adopted by VFX studios, architectural visualization firms, and small game studios that want exportable assets they can edit in DCC tools like Blender or Maya. Genie 3 is being explored by AI research labs, robotaxi companies, and developers who want a substrate for agent training rather than a content asset. As of early 2026 there is no direct user-base overlap.

Wayve GAIA-1 and GAIA-2

UK autonomous driving company Wayve introduced GAIA-1 in June 2023 and GAIA-2 in March 2025.^[21][35] Both are world models specialized for driving simulation, trained on Wayve's fleet data and supporting forecast of multi-camera driving video conditioned on action inputs.^[21] GAIA-2 in particular generates synchronized multi-camera output, which is essential for training the perception stack of a self-driving car.

The GAIA models are best understood as ancestors and parallel-track alternatives to Genie 3 rather than competitors. They are narrower (driving-only) but more useful for their specific domain, generating consistent driving scenes that integrate cleanly with autonomous vehicle development pipelines. Waymo's later collaboration with Genie 3 can be read partly as Waymo wanting Wayve-style capability without going through Wayve, and partly as a recognition that a general-purpose base model fine-tuned on driving data can outperform a specialist trained from scratch.

GAIA's significance to the world model space is foundational. Wayve's 2023 work was an early demonstration of a driving world model learned from video and conditioned on action inputs.^[21]

NVIDIA Cosmos

NVIDIA Cosmos, announced at CES in January 2025, is a family of world foundation models specifically targeted at "physical AI": robotics, autonomous vehicles, and industrial systems.^[22] Cosmos was released as a set of open-weight models that developers can fine-tune and integrate into their pipelines, contrasting with Genie 3's closed, hosted offering.^[22]

Cosmos and Genie 3 are arguably the closest direct philosophical competitors in the world model space, both aiming at "physical AI" use cases at a foundation model scale. The key differences are accessibility (Cosmos is open weights, Genie 3 is hosted), output framing (Cosmos provides simulation outputs designed to be consumed by training pipelines, Genie 3 provides interactive video designed to be consumed by humans or agents), and partner ecosystem (Cosmos has integration with NVIDIA's Omniverse and Isaac platforms, Genie 3 sits alongside SIMA and other Google systems).

For developers, the choice often comes down to deployment requirements. Researchers who need on-premises control over the model typically pick Cosmos. Researchers who want the highest available fidelity and breadth, and are comfortable with a hosted API or proprietary access, lean toward Genie 3.

Odyssey-1

Odyssey, a startup founded by autonomous vehicle researchers Oliver Cameron and Jeff Hawke, released Odyssey-1 as a playable world model with particularly low latency.^[12] The system generates video frames every 40 milliseconds from clusters of NVIDIA H100 GPUs, streaming interactive video at up to 30 frames per second.^[12] Odyssey described their approach as "interactive video" rather than 3D world simulation.^[12] At launch the environments were blurry and unstable, with layouts sometimes changing when the user turned around, but the streaming architecture and latency characteristics were notable.^[12]

Odyssey-1's framing as "interactive video" rather than "world simulator" is informative. The team explicitly positioned the product as a successor to passive video, not as a successor to game engines.^[12] Their target use case was novel forms of content (interactive films, exploratory media) rather than agent training. That framing has influenced how some observers think about Genie 3 as well, although DeepMind has consistently positioned Genie 3 closer to the agent-training pole.

Tencent HY World 1.5 and other open releases

Tencent's Hunyuan team released HY World 1.5 (WorldPlay) in December 2025 as an open-source, real-time interactive world model that streams video at 24 FPS.^[37] As a research release it lacks Project Genie's interface and tooling but is accessible to developers who want to build on top of a base model.

Other open-weight world model releases through 2025 and 2026 included community efforts based on Stable Video Diffusion variants and academic releases of smaller specialist models. None matched Genie 3's combination of fidelity, breadth, and interactivity in early 2026, but the open ecosystem is moving quickly, and the gap between hosted commercial models and downloadable alternatives is expected to narrow.

How was Genie 3 received?

Academic and technical reception to Genie 3 was largely positive on capability but mixed on transparency. The Hacker News discussion following the August 2025 announcement drew about 500 comments.^[11] One top commenter wrote, "Consistency over multiple minutes and it runs in real time at 720p? I did not expect world models to be this good yet," and added that world models "may have a bigger part to play in robotics and real world AI" than he had realized.^[11]

Criticisms focused on two areas. First, the announcement lacked an accompanying technical paper or detailed architecture description, which was unusual by DeepMind's standards.^[11] Multiple researchers and engineers expressed frustration that "value extraction" appeared to have taken precedence over academic transparency.^[11] DeepMind did not publish a preprint in conjunction with the announcement, leaving the community to infer architectural details from a blog post and a model page.^[1][2]

Second, a tester in the research preview, Tejas Kulkarni, wrote that physics showed "obvious failure cases," that social and multi-agent interactions were "tricky to handle," and that "long instruction following and simple combinatorial game logic fails."^[11] These reports aligned with DeepMind's own stated limitations but underscored the gap between the polished demo videos and typical session quality.

Media coverage was largely enthusiastic. TechCrunch headlined its report "DeepMind thinks its new Genie 3 world model presents a stepping stone toward AGI," a framing that itself drew debate.^[4] Critics noted that "AGI" is not a well-defined target and that positioning a video generation model as a stepping stone to general intelligence was a stretch that said more about marketing priorities than about the actual research contribution.

Game company shares fell on Friday, January 30, 2026, the day after the Project Genie launch. According to Reuters, Unity Software was down 21%, Roblox 12% and Take-Two Interactive 10% in that day's trading.^[30][31] Unity chief executive Matt Bromberg responded on X and LinkedIn that Project Genie's outputs were "unsuitable on their own for games that require consistent, repeatable player experiences". Games analyst Joost van Dreunen called the sell-off "an understandable mistake from Wall Street investors to paint the entire sector with the same brush" and argued that content, not infrastructure, determines winners in games.^[32]

How does Genie 3 fit into the AGI and embodied AI roadmap?

Genie 3 sits at a specific point in several converging AI research debates about the path to general intelligence.

The most influential of these is the position associated with Yann LeCun, who has argued for years that current language models lack the kind of world understanding that humans develop through embodied experience.^[23] In LeCun's framing, learned world models that predict how the world responds to actions are a prerequisite for systems that can plan, reason, and act in physical environments.^[23] Genie 3 fits this picture: it is a learned world model that responds to actions in real time, exactly the kind of capability LeCun has called for.

Demis Hassabis has consistently positioned DeepMind's research program around the idea that AGI will emerge from a combination of language reasoning, embodied learning, and search or planning. Genie 3 is the embodied-learning piece of that program at scale. Hassabis's public statements have repeatedly tied Genie 3 to a vision in which agents are trained across an unlimited variety of generated environments, eventually reaching a level of behavioral flexibility comparable to humans.

A third influence is David Silver and Richard Sutton's essay "Welcome to the Era of Experience," published in 2025, which argues that the next leap in AI capability will come from systems that learn primarily from their own experience rather than from human-labeled data.^[24] World models like Genie 3 are essential infrastructure for that approach: they provide the experience streams that agents act in and learn from at scales unreachable through real-world data collection alone.

Not all researchers accept this framing. Skeptics argue that pixel-level world models are too computationally expensive, too physically inaccurate, and too data-hungry to be the foundation of AGI. They suggest that more abstract or symbolic representations may be necessary, or that language-based reasoning systems combined with narrower physical simulators will outperform pure pixel-prediction approaches. The empirical question remains open as of early 2026.

What is less controversial is that Genie 3 has shifted the conversation. Before its release, world models were seen as a long-term research direction with no clear product path. After its release, multiple labs began publicly discussing world models as immediate priorities. Whether or not Genie 3 itself proves to be a stepping stone to AGI, it has at minimum established that real-time interactive world generation is technically feasible at meaningful fidelity, which has implications for the trajectory of the field whatever the eventual destination.

What are the limitations of Genie 3?

Google DeepMind has been explicit about the current limitations of Genie 3.^[1]

Session duration: Interactions are capped at a few minutes of continuous generation.^[1] The one-minute memory window and the growing trajectory computation cost mean that sessions cannot currently extend to hours, which would be needed for robust agent training curricula. Project Genie's 60-second cap at public launch reflected this practical ceiling.^[7]

Action space: The range of actions an agent can perform directly is constrained, and promptable world events are not necessarily performed by the agent itself.^[1] The range of controllable actions is narrower than hand-authored game environments.

Multi-agent scenarios: Genie 3 struggles when multiple independent agents share a generated space.^[1] Characters introduced through promptable world events do not behave as fully autonomous independent agents; they appear and move plausibly but do not maintain independent goals or memory.

Geographic accuracy: The model cannot reliably recreate specific real-world locations.^[1] Asking for "the plaza outside the Louvre" will produce something plausibly Parisian but not accurate to the actual site. This limits uses in architecture, urban planning, and geography education.

Text rendering: Text that appears within generated worlds, such as signs, labels, or written content, is typically illegible or distorted.^[1] This is a common failure mode for video generation models trained primarily on visual rather than text-layout data.

Text-only input at launch: At its August 2025 announcement, Genie 3 accepted only text prompts, unlike Genie 2, which could bootstrap from a single image.^[5] Image prompting was available by November 2025 and in Project Genie.^[6][10]

Physics accuracy: The learned physics simulation is approximate.^[1] Fluid dynamics, cloth behavior, and complex multi-body contacts are not reliably simulated.^[5] The model produces plausible-looking physics for most casual interactions but fails on scenarios requiring precise physical accuracy.

Resolution ceiling: 720p output is high enough for casual viewing but lower than current standards for VR, film production, or photoreal training pipelines. Upscaling to higher resolutions is possible in post-processing but does not recover the detail that a natively higher-resolution model would generate.

Compute cost: Real-time autoregressive video generation at this fidelity is expensive. The unit economics of running Genie 3 sessions at consumer scale are unclear, and there is no public information about whether Google can serve the model profitably at Google AI Ultra prices ($249.99 a month at launch; Google described the eligible tier as "Google AI Ultra $200" in May 2026)^[33] or whether the launch is being subsidized as a research investment.^[6][29]

Reproducibility: The probabilistic nature of generation means the same prompt and the same actions produce different sessions on different runs. For agent training that depends on repeatable benchmarks, this is a significant constraint. Workarounds, such as seeding the generator and replaying inputs, are imperfect because small numerical differences in compute can shift outcomes.

Safety and responsibility: Google DeepMind's responsible AI documentation for Genie 3 notes that the open-ended and real-time nature of the system creates novel safety challenges compared to text or image generation.^[1] The limited research preview rollout before Project Genie reflected caution about unforeseen uses at scale.^[1] Specific concerns identified include the potential for generating violent or disturbing content from innocuous prompts, the use of the system to author misleading visual narratives, and the difficulty of moderating content generated dynamically rather than uploaded as static media.

Limited public access initially: Even after Project Genie launched, access was confined to the US, to adult users, and to the $249.99 Google AI Ultra tier.^[6][29] International expansion and lower-cost access tiers were promised but not scheduled at launch; the worldwide rollout to AI Ultra subscribers followed in May 2026, though access remained tied to the top-tier subscription.^[6][26] This limited access has constrained external research on the system's properties and failure modes, and has been a recurring criticism in academic discussion.

Continue reading

• Genie (DeepMind)Genie is an 11-billion-parameter generative interactive environment from Google DeepMind, described by its creators as the first foundation world model: it turns a single image, photo, or sketch into a…Mentioned 177 times above
• WWorld modelA world model is an artificial intelligence system that learns an internal representation of how an environment works, enabling it to predict future states, simulate the consequences of actions, and support…Mentioned 38 times above
• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 26 times above
• Genie 2Genie 2 is a foundation world model developed by Google DeepMind, unveiled on December 4, 2024.Mentioned 14 times above
• NVIDIA CosmosNVIDIA Cosmos is a world foundation model platform developed by NVIDIA for physical AI applications, including autonomous vehicles and robotics.Mentioned 9 times above
• WayveWayve is a British artificial intelligence company, founded in Cambridge in 2017 by Alex Kendall and Amar Shah, that develops end to end "embodied AI" software for autonomous driving and robotics, and in May…Mentioned 8 times above

More to read

• SIMA (DeepMind)
• Genie 4
• Sora
• World Labs
• Veo 3
• Embodied AI

Topics:AI ModelsGenerative AIGoogle DeepMindWorld Models

See also

• World model
• Google DeepMind
• Genie
• Genie 2
• Genie 4
• Sora
• Veo 3
• Embodied AI
• SIMA
• World Labs
• Oasis (model)
• Microsoft Muse
• NVIDIA Cosmos
• Wayve
• Generative AI
• NeRF
• Gaussian splatting

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34Google DeepMind. "Genie 3: A new frontier for world models." *Google DeepMind Blog*, August 5, 2025. deepmind.google/...a-new-frontier-for-world-models
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google DeepMind. "Genie 3." *Google DeepMind Models*, 2025. Archived 25 October 2026: web.archive.org/...genie (the live address, deepmind.google/...genie, has presented Genie 4 since 27 October 2026).
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., et al. "Genie: Generative Interactive Environments." *arXiv*, February 23, 2024. arxiv.org/...2402.15391
4. ^1 ^2 ^3 ^4Bellan, R. "DeepMind thinks its new Genie 3 world model presents a stepping stone toward AGI." *TechCrunch*, August 5, 2025. techcrunch.com/...sents-stepping-stone-towards-agi
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Dickson, B. "The promise and limitations of DeepMind's Genie 3." *TechTalks*, August 7, 2025. bdtechtalks.com/...deepmind-genie-3
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Google. "Project Genie: AI world model now available for Ultra users in U.S." *Google Blog*, January 29, 2026. blog.google/...project-genie
7. ^1 ^2 ^3 ^4 ^5Bellan, R. "I built marshmallow castles in Google's new AI-world generator, Project Genie." *TechCrunch*, January 29, 2026. techcrunch.com/...ai-world-generator-project-genie
8. ^Li, A. "Google rolling out 'Project Genie' to generate playable worlds." *9to5Google*, January 29, 2026. 9to5google.com/...google-project-genie
9. ^1 ^2 ^3 ^4 ^5Waymo. "The Waymo World Model: A New Frontier For Autonomous Driving Simulation." *Waymo Blog*, February 6, 2026. waymo.com/...ier-for-autonomous-driving-simulation
10. ^1 ^2 ^3 ^4Google DeepMind. "SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds." *Google DeepMind Blog*, 2025. deepmind.google/...s-with-you-in-virtual-3d-worlds
11. ^1 ^2 ^3 ^4 ^5 ^6Hacker News discussion. "Genie 3: A new frontier for world models." *Hacker News*, August 5, 2025. news.ycombinator.com/item
12. ^1 ^2 ^3 ^4 ^5Wiggers, K. "Odyssey's new AI model streams 3D interactive worlds." *TechCrunch*, May 28, 2025. techcrunch.com/...el-streams-3d-interactive-worlds
13. ^1 ^2 ^3Statt, N. "Fei-Fei Li's World Labs speeds up the world model race with Marble." *TechCrunch*, November 12, 2025. techcrunch.com/...ble-its-first-commercial-product
14. ^Togelius, J. "Genie 3 and the future of neural game engines." *Julian Togelius Blog*, August 2025. togelius.blogspot.com/...and-future-of-neural-game
15. ^1 ^2 ^3 ^4"Genie (world model)." *Wikipedia*, 2026. en.wikipedia.org/...Genie_(world_model)
16. ^"World Models Race 2026." *Introl Blog*, 2026. introl.com/...world-models-race-agi-2026
17. ^"Waymo Says Genie 3 Simulations Can Help Boost Robotaxi Rollout." *Bloomberg*, February 6, 2026. bloomberg.com/...s-can-help-boost-robotaxi-rollout
18. ^1 ^2 ^3 ^4Microsoft Research. "Introducing Muse: Our first generative AI model designed for gameplay ideation." *Microsoft Research Blog*, February 19, 2025. microsoft.com/...el-designed-for-gameplay-ideation
19. ^1 ^2Kanervisto, A., et al. "World and Human Action Models towards gameplay ideation." *Nature*, February 2025.
20. ^1 ^2Harper, C. "AI-powered Minecraft runs without a game engine: game rendered in real-time at a continuous 20 FPS." *Tom's Hardware*, November 2, 2024. tomshardware.com/...al-time-at-a-continuous-20-fps
21. ^1 ^2 ^3Wayve. "Scaling GAIA-1: 9-billion parameter generative world model for autonomous driving." *Wayve Blog*, October 3, 2023. wayve.ai/...scaling-gaia-1
22. ^1 ^2NVIDIA. "NVIDIA Cosmos World Foundation Model Platform." *NVIDIA Blog*, January 2025 (CES). nvidia.com/...cosmos
23. ^1 ^2LeCun, Y. "A Path Towards Autonomous Machine Intelligence." *OpenReview*, 2022.
24. ^Silver, D. and Sutton, R. S. "Welcome to the Era of Experience." 2025. storage.googleapis.com/...20Experience%20Paper.pdf
25. "Genie 3 (world model)." *Wired*, August 2025.
26. ^1 ^2"Project Genie (website)." *Wikipedia*, 2026. en.wikipedia.org/...Project_Genie_(website)
27. ^1 ^2 ^3Hadsell, R., Rivas, D., and Qi, H. "Genie 4: A new generation of world simulation." *Google DeepMind Blog*, October 27, 2026. deepmind.google/...-generation-of-world-simulation
28. ^Giret, L. "Google Makes Genie 4 World Simulation Model Available for Select Partners." *Thurrott.com*, October 27, 2026. thurrott.com/...odel-available-for-select-partners
29. ^1 ^2 ^3 ^4Wiggers, K. "Google AI Ultra: You'll have to pay $249.99 per month for Google's best AI." *TechCrunch*, May 20, 2025. techcrunch.com/...99-per-month-for-googles-best-ai
30. ^Bezmalinovic, T. "Google's new AI tool causes gaming stocks to plummet." *heise online*, February 2, 2026. heise.de/...uses-gaming-stocks-to-plummet-11162226
31. ^"Immediately after Google announced the global model 'Project Genie,' stock prices of game companies fell across the board." *GIGAZINE*, January 31, 2026. gigazine.net/...-stocks-slide-google-project-genie
32. ^Long, N. "Unity boss Bromberg moves to calm markets as Google's Genie causes gaming stock slide." *Mobilegamer.biz*, February 2, 2026. mobilegamer.biz/...genie-causes-gaming-stock-slide
33. ^1 ^2 ^3Rivas, D. and Herbert, J. "Simulate real-world places with Project Genie and Street View." *Google Blog*, May 19, 2026. blog.google/...project-genie-expands
34. ^1 ^2Google DeepMind. "Genie 2: A large-scale foundation world model." *Google DeepMind Blog*, December 4, 2024. deepmind.google/...ge-scale-foundation-world-model
35. ^Russell, L., et al. "GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving." *arXiv*, March 26, 2025. arxiv.org/...2503.20523
36. ^"GDC Survey: Over 50% Of Game Devs Say Generative AI Harms Industry." *80 Level*, January 30, 2026. 80.lv/...ame-devs-say-generative-ai-harms-industry
37. ^1 ^2Tencent Hunyuan. "HY-World 1.5 (WorldPlay)." *GitHub*, December 2025. github.com/...HY-WorldPlay
38. ^1 ^2Wiggers, K. "Decart's AI simulates a real-time, playable version of Minecraft." *TechCrunch*, October 31, 2024. techcrunch.com/...me-playable-version-of-minecraft

Show all 38 references