World models and games, what Genie 3 and Oasis can and cannot do

Rick Vu, 9 October 2026

An AI agent drafted this piece under Rick's direction, as part of an experiment in AI writing. Every figure links to its source.

A world model in games is an AI model that generates the next frame of an interactive world from the frames before it and the player’s latest input. There is no game engine, no level file, and no stored game state behind the picture. Google DeepMind’s Genie 3, Decart’s Oasis, and Microsoft’s Muse are the best-known examples, and each keeps its world consistent for seconds or minutes, not for a full play session.

This piece is dated October 9, 2026. It explains how a world model works, what three labs have shown, and what their own posts say the models cannot do yet. For where world-model labs sit among other AI gaming startups, and for Decart’s funding history, see the field map of AI-native startups in gaming. This piece stays on the technology and on what a studio could do with it today.

How a world model makes a frame

A game engine keeps the world as data. It knows where every wall is, how much health the player has, and which door is locked. Each frame, the engine updates that data from the player’s input, and a renderer draws the picture from it. The picture is a view of the state.

A world model has no such data. It learns from video of play paired with the inputs that produced it. Decart’s Oasis announcement says the model takes keyboard input and generates frames one at a time, with each frame conditioned on that input. Microsoft trained Muse on more than 1 billion images and controller actions from the Xbox game Bleeding Edge, which Microsoft says equals over seven years of continuous human play.

At run time the loop is short. The model looks at recent frames and the key the player pressed, predicts the next frame, adds it to the history, and repeats. Google DeepMind’s Genie 3 announcement says the model has to take into account a trajectory that grows with time, and has to do that computation multiple times per second as new input arrives. Any rule the player sees, such as gravity or a door that swings open, exists only as a pattern the model learned from footage.

What has shipped or been shown

Three labs have shown world models you can control in real time. They built them for different jobs.

Genie 3 and Project Genie

Google DeepMind announced Genie 3 on August 5, 2025. The Genie 3 announcement says you can move through its worlds in real time at 24 frames per second and 720p. It says the worlds stay consistent for a few minutes and that visual memory reaches back about one minute. You can also change a running world with a text prompt, for example to change the weather or add a character. DeepMind first released it as a limited research preview for a small group of academics and creators.

The version the public can touch came later. Google’s Project Genie announcement of January 29, 2026, describes a prototype web app powered by Genie 3, Nano Banana Pro, and Gemini. It opened to Google AI Ultra subscribers in the US aged 18 and over, and it limits each generation to 60 seconds. Pulse 2.0’s report on the Project Genie expansion in May 2026 says access is expanding to eligible Google AI Ultra subscribers worldwide.

Oasis from Decart

Decart and the chip company Etched released Oasis on October 31, 2024, as a Minecraft-like world generated frame by frame. The Oasis announcement claims real-time output at 20 frames per second. TechCrunch’s report on the Oasis demo found the resolution “quite low” and said the model tends to “forget” the level layout, so turning around can show a rearranged landscape. TechCrunch also noted that Decart did not say it had Microsoft’s blessing to train on Minecraft footage.

Oasis has since changed role. The Next Web’s report on Decart’s $300 million round in May 2026 describes Oasis as Decart’s product for physical AI, aimed at robotics and autonomous systems. Gaming now sits as a use case under Oasis and under Decart’s Lucy model.

Muse and WHAMM from Microsoft

Microsoft published Muse in the journal Nature in February 2025. The Muse announcement calls it a World and Human Action Model, or WHAM, and frames it as a tool for gameplay ideation rather than a game. Its output is 300 by 180 pixels, and Microsoft says it generates consistent gameplay sequences of up to two minutes. It can also carry a change a designer makes, such as an added character, into later frames. Microsoft open-sourced the weights and a demonstrator app.

In April 2025 Microsoft put a playable follow-up in Copilot Labs. The WHAMM research article describes a model that plays part of one Quake II level at more than 10 frames per second and 640 by 360 pixels. Microsoft trained it on one week of data.

What the published limits are

The table collects the figures each lab published. Where a cell says “not stated”, the lab’s own post gives no number.

The posts list limits beyond the numbers. The Genie 3 announcement says the range of actions an agent can take is limited, that interaction between several independent agents is an open research problem, and that legible text usually appears only when the prompt includes it. The Project Genie announcement adds that worlds may not follow prompts closely, may break real-world physics, and may give characters higher control latency. The WHAMM article says enemies look fuzzy, combat results can be wrong, and the health value “is not always super reliable.”

None of these posts gives a cost per second of play. TechCrunch reported that the Oasis demo ran on Nvidia H100 GPUs, which are data-center accelerators. Here is an estimate, not a published figure. Until a model like this runs on the player’s own hardware, every minute of play is a minute of server GPU time that someone pays for. A single-player engine game runs on the player’s machine, so an extra hour of play costs the studio no compute.

Why world models forget what is behind the player

A world model only knows what is in its context, the recent frames it reads before it predicts the next one. When the player turns away from a wall, the wall stays in the context for a while. Once those frames drop out, nothing records that the wall existed. The model predicts a view that looks plausible, and plausible is not the same as the one the player saw.

The window sizes vary a lot. The WHAMM article puts its context at 0.9 seconds, which is nine frames at 10 frames per second, and says the model can forget objects that leave the view. Genie 3’s visual memory reaches back about a minute. Both are far shorter than the gap before a player walks back into a level they left twenty minutes ago.

DeepMind says Genie 3’s consistency is an emergent capability. Its post contrasts the model with methods such as NeRFs and Gaussian splatting, which depend on an explicit 3D representation of the scene. That is the trade. With no stored geometry, the model can make up any world it can describe. It also has nowhere to keep one.

The same gap covers state the player cannot see. An inventory, a quest flag, or a score lives in an engine as a variable. In a world model it has to survive as pixels on screen, and Microsoft’s note about the WHAMM health value is one visible case. A longer memory helps, but the Genie 3 post says the trajectory grows with time, so each extra second of memory adds work to every frame.

Will world models replace game engines

Not on the evidence published so far. A shipped game needs rules that hold for hours, and these models hold for minutes at best. A save file has to load the same world tomorrow. In multiplayer, every player has to see the same world, and DeepMind lists interaction between several agents as an open problem. A team also has to fix a bug in one place and know it stays fixed. A learned model gives the designer no line of code to change, only more training data or a different prompt.

These are limits of the current generation, and the published figures moved fast between late 2024 and 2025. Oasis showed 20 frames per second of a blurry Minecraft-like world, and less than a year later Genie 3 showed 720p worlds that stay consistent for minutes. A studio should plan around the models it can use today and check the limits again when the next generation ships.

Where a world model could sit next to an engine

The useful question for a studio today is what a world model can do beside an engine. Each idea below is a proposal sized to the limits above, not a report of a studio doing it.

• Pitches and mood sketches. A 60-second playable world from a prompt or a concept image can show a team how a place might feel before anyone builds a greybox level. Project Genie lets you download a video of the world and your path through it.
• Gameplay ideation. Microsoft built Muse for this. A designer adds a character or an item to a frame and looks at how play might branch from there.
• Testing agents. DeepMind ran its SIMA agent inside Genie 3 worlds, where the agent sent navigation actions while it pursued goals. A studio that trains bots could use generated worlds for variety, then test the bots in the real game.
• Short scenes where drift fits. A dream sequence or a vision that lasts under a minute can live with a world that shifts, because the player expects it to shift. Nobody publishes a cost per second, so price it before you design it in.

In each case the engine keeps the state that matters, and the world model makes pictures around it.

Related reading

• AI-native startups in the gaming market in 2026, for where world-model labs sit among the other bets.
• AI game maker or game engine, which one you need, for the other AI tool people expect to replace engines.
• How AI NPCs remember, for the same memory limit in a character instead of a world.