Genie 3
Last reviewed
Jun 5, 2026
Sources
25 citations
Review status
Source-backed
Revision
v4 ยท 8,612 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
25 citations
Review status
Source-backed
Revision
v4 ยท 8,612 words
Add missing citations, update stale details, or suggest a clearer explanation.
Genie 3 is a foundation world model developed by Google DeepMind, announced on August 5, 2025.[1] It generates photorealistic, interactive three-dimensional environments from text descriptions, runs in real time at 24 frames per second and 720p resolution, and maintains environmental consistency for several minutes of continuous interaction.[1] Google DeepMind describes Genie 3 as "a crucial stepping stone on the path to artificial general intelligence," positioning it primarily as a training substrate for general-purpose AI agents.[1]
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] The system taught itself basic physics entirely from video data, without any hard-coded physics engine.[1]
Genie 3 became publicly accessible through Project Genie on January 29, 2026, initially limited to Google AI Ultra subscribers in the United States.[6] The release marked the first time a major lab put a real-time interactive world model into the hands of consumers, even at an experimental tier.
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.
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.
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. Ars Technica reported at the time that DeepMind's claim of roughly one minute of consistency was optimistic for most use cases. 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 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, described the release in posts on X as "a new frontier for world models" 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 through the fall, gathering feedback from selected academic and industry partners.[1] Project Genie, the consumer-facing launch, did not arrive until January 29, 2026.[6]
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. Shlomi Fruchter of Google DeepMind 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.
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 represents a roughly six-fold improvement over Genie 2's effective limit of 10 to 20 seconds, 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.
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.[1]
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.
In DeepMind's research preview, this feature was the most cited differentiator versus both Decart Oasis and World Labs Marble. 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.
Genie 3 was not given a physics engine during training. Its understanding of how objects fall, how water behaves, how lighting changes with weather, and how terrain deforms came entirely from watching video data.[1] 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. DeepMind's own blog acknowledged inaccuracies, giving the example of snow movement around a skier showing errors.[1] 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.
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 specifying art directions like "watercolor painting" or "in the style of Studio Ghibli."[1]
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.
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:
None of these inferences have been confirmed in detail by DeepMind. The blog post mentions "a number of new advances," 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.
Genie 3 does not accept image inputs for world initialization; as of its August 2025 announcement, the model takes only text prompts.[1] 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]
Independent estimates from researchers at competing labs suggest Genie 3 likely requires multiple TPU v5p or H100 GPUs per concurrent user, depending on how the autoregressive context is sharded across accelerators. The economics of running such a system at consumer scale are unclear; the Project Genie tier through Google AI Ultra at $249.99 per month implies Google is willing to absorb significant per-session compute costs to gather usage data and feedback.[6]
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.
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 costs $249.99 per month, making Project Genie accessible primarily to developers, researchers, and enthusiasts already invested in Google's AI ecosystem.[6]
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.[6] 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] Google acknowledged this limit was a product decision rather than a technical ceiling, citing the experimental nature of the prototype.[6] 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 expansion to other regions would happen "in the future" without specifying a timeline.[6]
Reaction to Project Genie's public availability was mixed in tone but enthusiastic in interest. TechCrunch's Devin Coldewey published a hands-on piece titled "I built marshmallow castles in Google's new AI-world generator," describing the experience as alternately magical and bewildering.[7] He wrote that the sessions felt like "a memory of a place you have never been," and noted that the 60-second cap kept any single experience from outstaying its welcome.[7] 9to5Google's Jess Elias took a more pragmatic angle, describing the use cases for game developers, educators, and worldbuilders, and noting that the gallery and remixing features suggested Google was treating Project Genie as a social platform in addition to a research showcase.[8]
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, released in late 2025, showed that agents trained partly in Genie 3-generated worlds could transfer their skills to real game environments they had never seen during training.[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.
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, at four times the lidar output speed of base Genie 3.[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 represents the first publicly announced industry application of Genie 3 outside Google 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 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.
A researcher at Game Developer Conference 2026 found that 52% of gaming professionals viewed AI tools as having a negative impact on the industry (up from 30% in 2025 and 18% in 2024), 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, observed that Genie 3 represents a step toward what he called a "neural game engine" but argued that actual games require far more than consistent world generation.[14] Precise collision detection, deterministic state, and author-controlled pacing are properties that pixel-level prediction systems cannot yet reliably provide.[14]
Google positions Genie 3 for rapid prototyping and concept exploration, not as a replacement for commercial game development pipelines. Studios that have experimented with the system, including several anonymized partners in DeepMind's research preview, reported using it for early concept work where the goal is to communicate a feel rather than implement a final design. 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.
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.
A few examples of creative applications surfaced in the months after Project Genie's launch. Two short experimental films were released that used Genie 3 footage as their primary visual material, both running under five minutes and explicitly framed as art pieces rather than narratives. A handful of musicians used generated worlds as music video backdrops. Some VR companies began experimenting with rendering Genie 3 output to headset displays, though motion sickness from the model's occasional visual artifacts was an early issue.
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.
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.
| Model | Developer | Release date | Output format | Resolution/FPS | Interactivity | Key differentiator |
|---|---|---|---|---|---|---|
| Genie | Google DeepMind | February 23, 2024 | Pre-rendered 2D video | 160x90 @ 1 FPS | No real-time | First foundation world model from video |
| Genie 2 | Google DeepMind | December 4, 2024 | Generated 3D video | 360p @ slow | Limited | 3D upgrade from Genie 1, 10-20s memory |
| Genie 3 | Google DeepMind | August 5, 2025 | Real-time navigable video | 720p @ 24 FPS | Yes, keyboard-controlled | Long memory window, promptable events |
| Oasis | Decart AI / Etched | October 31, 2024 | Real-time video | ~480p @ 20 FPS | Yes | Browser-accessible, Minecraft-style |
| Muse | Microsoft Research | February 19, 2025 | Game state and pixel prediction | Game-dependent | Yes | Bleeding Edge game model |
| Odyssey-1 | Odyssey | May 28, 2025 | Streaming video | 480p @ 30 FPS | Yes | 40ms latency, interactive video |
| Marble | World Labs | November 12, 2025 | 3D scene assets | Exportable 3D | Limited | Downloadable, editable 3D output |
| HY World 1.5 | Tencent Hunyuan | Late 2025 | Real-time video | 720p @ 24 FPS | Yes | Open source |
| GWM-1 Worlds | Runway | Late 2025 | Real-time navigable video | 720p real-time | Yes | Spatial coherence on revisit |
| GAIA-1 | Wayve | April 2023 | Driving video | 480p | Limited | Driving-specific world model |
| GAIA-2 | Wayve | December 2024 | Multi-camera driving video | Multi-stream | Limited | Multi-camera driving |
| Cosmos | NVIDIA | January 2025 (CES) | Physical AI WFMs | Various | Limited | Open weights for physical AI |
Decart AI released Oasis in late October 2024, in collaboration with hardware startup Etched.[20] 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 designed from the start to run on custom inference hardware (Etched's specialized transformer chip), which made the economics of real-time playable generation more favorable than running on general-purpose GPUs.[20] 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 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.
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.
UK autonomous driving company Wayve released GAIA-1 in April 2023 and GAIA-2 in December 2024.[21] 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 paper was among the first to demonstrate that driving simulators could be learned end-to-end from video, and its approach of conditioning generation on explicit action inputs influenced later systems including Genie 1 and Genie 3.[21]
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, 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's Hunyuan team released HY World 1.5 as an open-source world model running at 720p and 24 FPS, matching Genie 3's specifications on paper and providing the first fully open-source option at that resolution. 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.
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 over a thousand comments.[11] Several developers who tested the model expressed surprise that it achieved "consistency over multiple minutes AND runs in real time at 720p" simultaneously, calling it an advance beyond their expectations.[11] Robotics researchers were particularly interested, with multiple commenters observing that world models could have "a bigger part to play in robotics and real world AI" than previously assumed.[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, developers who tested the limited research preview reported that physics failures were common, social interaction scenarios did not work, and complex instruction following "fails in surprising and obvious ways."[5] One developer noted that long instruction sequences and simple combinatorial game logic were both unreliable. 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 described it as DeepMind's "stepping stone toward AGI" framing, which 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.
Wired ran a feature focused on Genie 3's implications for content creation and the games industry, exploring whether the system could displace concept artists and level designers.[25] The piece concluded that displacement risk was lower than initial reactions suggested because the model could not yet produce ship-ready content, but acknowledged that the workflow disruption was real even at the ideation stage.[25]
MIT Technology Review approached the announcement from the AGI angle, treating Genie 3 as evidence that world models were emerging as a credible third pillar of foundation AI alongside language and vision models. The article quoted multiple researchers, including some skeptical voices, arguing that learned simulators are essential infrastructure for the kind of embodied learning that AGI would require.
Demis Hassabis, Google DeepMind's CEO, was unusually active in promoting Genie 3 on social media. His posts on X repeatedly framed the model as part of DeepMind's roadmap toward AGI through embodied learning, drawing parallels with positions held by Yann LeCun and Rich Sutton about the importance of experience and simulation for general intelligence. The framing prompted skeptical responses from researchers who argued that DeepMind was overselling the model's research significance.
The stock market reacted to the Project Genie public launch in January 2026. Unity Software fell 21%, Roblox Corporation fell 15%, Take-Two Interactive fell 9.3%, and CD Projekt fell 8% in the days following the announcement, reflecting investor concern about Genie 3's potential to displace conventional game development tools. Analysts disagreed about whether the reaction was warranted. Some argued that the existing game companies' moats around content, IP, and live operations were intact regardless of generation tooling; others noted that the cost dynamics of game production would shift if competitive prototypes could be built without large content teams.
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 Rich Sutton's "Era of Experience" framing, published in 2024 and 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.
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: User control is limited to movement navigation.[1] There is no support for fine manipulation, object grasping, or other actions relevant to robotic or embodied agent training beyond locomotion.[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: As of its August 2025 announcement, Genie 3 accepts only text prompts as input, unlike Genie 2 which could bootstrap from a single image.[1] This restriction limits the ability to anchor generated worlds to a specific visual reference.
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 the $249.99 Google AI Ultra price point or whether the launch is being subsidized as a research investment.[6]
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] International expansion and lower-cost access tiers were promised but not scheduled.[6] This limited access has constrained external research on the system's properties and failure modes, and has been a recurring criticism in academic discussion.