Every major AI lab spent 2023 through 2025 racing to build a bigger chatbot. In 2026, the race quietly shifted. Google DeepMind, Fei-Fei Li’s World Labs, NVIDIA, Tencent, and the real-time video startup Decart are now competing to build something different: AI models that don’t just answer questions, but generate entire interactive 3D worlds you can walk through, drive in, or train a robot inside. Industry watchers call this category “world models,” and it’s the space Yann LeCun left Meta to chase full-time.
This comparison breaks down the three products drawing the most search interest right now — Google DeepMind’s Genie 3, World Labs’ Marble, and NVIDIA’s Cosmos platform — and brings in two more players, Decart’s Oasis/Lucy and Tencent’s HunyuanWorld line, because no world-model comparison is complete without them. We’ll cover specs, pricing, funding, real-world use cases, a migration path for teams evaluating these tools, and a data-backed verdict on which model fits which job.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is an AI World Model, and Why Does It Matter Now?
A world model is a generative AI system trained to simulate a coherent, spatially consistent environment rather than a single image or a block of text. Feed it a text prompt, a photo, or a video clip, and it outputs a 3D or video environment you can move through, manipulate, or extract as a usable asset. That’s a fundamentally different task than what GPT-5.6, Claude, or Gemini do. Those are language and reasoning engines. World models are spatial simulators.
Google framed the shift plainly when it described Project Genie as “experimenting with infinite, interactive worlds” in its DeepMind announcement. That’s a different pitch than “answer my question” — it’s “generate a place, and let me act inside it.” NVIDIA calls its equivalent category “world foundation models” (WFMs), and pitches them squarely at robotics and autonomous-vehicle teams that need synthetic training data no camera crew could ever capture in the real world.
The money backs up the hype. World Labs has raised roughly $1.23 billion since its founding, and Decart has pulled in more than $450 million across five disclosed rounds, most recently a $300 million round in May 2026 led by Radical Ventures with NVIDIA, Adobe Ventures, and Toyota Ventures all writing checks. When Toyota and NVIDIA both invest in the same world-model startup, it’s a signal that this category isn’t a research curiosity anymore — it’s being built into how carmakers and chipmakers plan to train physical AI systems for the rest of the decade.
World Models vs AI Video Generators: What’s the Difference?
It’s easy to lump Genie 3, Marble, and Cosmos in with text-to-video tools like Sora, Veo, or Kling, since all of them turn a prompt into moving pixels. The underlying architecture and end goal are different enough that treating them as the same category leads to the wrong buying decision.
A video generator produces a fixed, linear clip. You type a prompt, wait, and get back a video file with a beginning and an end that you can’t step into or redirect once generation starts. A world model produces a state you can act inside. Genie 3 lets you steer where you look and move in real time, not just watch a pre-rendered result. Marble goes a step further by keeping that state persistent — the 3D scene it builds doesn’t disappear when the session ends, it exports as a file you can reopen, edit, and walk through again in a completely different tool.
That distinction matters most for the robotics and simulation use cases NVIDIA Cosmos and Decart’s Oasis target. A video generator can produce a convincing-looking driving scene, but it has no concept of object permanence or physical consistency once the camera pans away and back. A world model has to maintain that consistency, because the entire point is to generate data (or an experience) a downstream system can act on without breaking. That’s a harder technical problem, which is part of why Cosmos leans on named sub-components like Cosmos Predict, Transfer, and Reason instead of shipping a single monolithic model the way most video generators do.
The practical rule of thumb: if you need a finished video clip for marketing, social content, or a film sequence, a text-to-video generator is still the right tool and generally faster and cheaper per output. If you need something a person, a robot, or a simulated vehicle can navigate, explore, or train against, you’re in world-model territory, and the five products in this comparison are the ones actually built for that job.
Why Investors and Big Tech Are Betting Big on World Models
The funding numbers in this category look unusual for how early the products still are. Most of these tools shipped their first real public version within the last 12 months, yet World Labs alone has raised more than a billion dollars, and Decart’s valuation quadrupled in the space of nine months, from roughly $1 billion pre-Series B territory to about $4 billion by its May 2026 round.
Part of the explanation is who’s writing the checks. Decart’s May 2026 round pulled in NVIDIA, Toyota Ventures, and Adobe Ventures alongside its existing lead investors Sequoia Capital and Benchmark. That’s not a typical growth-equity syndicate — it’s a chipmaker, an automaker, and a creative-software giant, each of whom has a direct operational reason to want cheaper, faster synthetic environment generation. NVIDIA needs world models to sell more Omniverse and robotics simulation licenses. Toyota needs them to cut the cost of autonomous-driving validation. Adobe needs them to keep its creative suite relevant as 3D content generation shifts from manual modeling to AI generation.
Yann LeCun’s move is the clearest signal of where the research community’s confidence is heading. LeCun, one of the three researchers who shared the 2018 Turing Award for deep learning, spent years at Meta arguing publicly that large language models were a technological dead end for building systems with real understanding of the physical world. He left to build a world-model-focused venture that has since raised north of $1 billion, a scale of funding usually reserved for labs building frontier LLMs, not a niche research bet. When a researcher with that level of credibility stakes his post-Meta career on world models specifically, and investors back that bet at a billion-dollar-plus scale, it tells you the field sees this as a genuine architectural fork in AI development, not a side quest.
None of this guarantees the current crop of products will be the ones still standing in two years. Genie 3’s 60-second memory limit, the lack of published benchmarks across the board, and the sheer number of well-funded competitors all point to a category that’s still working out its winning architecture. But the capital commitment from Google, NVIDIA, Tencent, and a growing list of independent-lab investors makes it very unlikely this category shrinks back down to a research curiosity anytime soon.
Genie 3 vs Marble vs NVIDIA Cosmos: Full Specs Table
Here’s how the five most relevant world models stack up on the specs that matter most to developers, studios, and robotics teams evaluating this category in September 2026.
| Spec | Google DeepMind Genie 3 | World Labs Marble | NVIDIA Cosmos (WFMs) | Decart Oasis / Lucy | Tencent HY-World 2.0 |
|---|---|---|---|---|---|
| Category | Real-time interactive world simulation | Persistent 3D world generation | Physics-based synthetic data / simulation | Real-time video world model | 3D world generation & reconstruction |
| Latest release | Genie 3 (Project Genie), gated rollout began Jan. 29, 2026 | Marble 1.1 / 1.1 Plus, April 2, 2026 | Cosmos 3 flagship, described in 2026 platform overview | Oasis 3, June 2026 | HY-World-2.0, April 16, 2026 |
| Output resolution / frame rate | ~720p at ~24 fps (reported) | Full-resolution 3D scenes, Gaussian splat export | Not publicly disclosed; focus is physics accuracy over resolution specs | Real-time, sub-30ms response reported for Lucy | Mesh or 3D Gaussian splat output, resolution not disclosed |
| Session / memory length | ~60 seconds of interactive coherence | Persistent (exportable, not session-limited) | N/A — generates standalone synthetic clips/data | Continuous real-time stream | Persistent, reconstructable scenes |
| Input types | Text prompt (within Genie interface) | Text, image, video, panorama | Text, image, video, robot sensor/motion data | Video stream, text | Text, single/multi-view images, video |
| Licensing | Closed, subscription-gated | Closed, freemium + paid tiers | Open model license (Hugging Face, NGC) | Closed, commercial API | Open-source, GitHub + Hugging Face |
| Access requirement | Google AI Ultra $200/mo tier, 18+, US only at launch | Free tier available; no subscription required to start | Free download under open model license | API access, credit-based | Free download, self-hosted |
| Pricing (paid tiers) | $200/month (Google AI Ultra) | Free / $20 Standard / $35 Pro / $95 Max per month | Free (compute costs apply for self-hosting) | API credits, rates not publicly listed | Free (compute costs apply for self-hosting) |
| Commercial usage rights | Included in Ultra subscription | Begins at Pro tier ($35/mo) | Permitted under open model license terms | Included in API terms | Per open-source license terms |
| Primary use case | Interactive world exploration, game-like prototyping | World building, VFX/previs, game prototyping, simulation | Robotics and AV synthetic training data | Gaming, e-commerce, advertising, AV driving simulation | VR/AR environments, game world creation |
| Backing / funding | Google (internal, DeepMind) | ~$1.23B raised total | NVIDIA (internal platform) | $450M+ raised, ~$4B valuation (May 2026) | Tencent (internal) |
A few things jump out immediately. Genie 3 is the most restricted on paper — it’s locked behind a $200-a-month consumer subscription, geo-limited to the US, and age-gated to 18 and older. NVIDIA Cosmos and Tencent’s HY-World line sit at the opposite end: both ship as downloadable, open-license weights that a developer can run on their own GPUs today. Marble splits the difference with a genuine free tier plus paid subscriptions that unlock commercial rights and higher output quality. Decart’s Oasis and Lucy are the odd ones out — commercial-only, API-gated, and priced on a credit system that isn’t published in flat per-month terms.
Pricing Breakdown: What Each World Model Actually Costs
Pricing across this category ranges from completely free to $200 a month, and the structures don’t map onto each other cleanly — some charge per subscription tier, some per API credit, and some don’t charge at all because you’re expected to bring your own compute.
| Model | Free tier? | Entry paid price | Top paid price | API access |
|---|---|---|---|---|
| Genie 3 | No | $200/mo (Google AI Ultra) | $200/mo (single tier with access) | Not publicly available |
| Marble | Yes | $20/mo (Standard) | $95/mo (Max) | Yes, $1 per 1,250 credits |
| NVIDIA Cosmos | Yes (open weights) | $0 (self-hosted) | Compute cost only | Preview access via NVIDIA API catalog + NIM microservices |
| Decart Oasis / Lucy | No | Credit-based (rates not public) | Credit-based (rates not public) | Yes, sold as API to AV developers |
| Tencent HY-World 2.0 | Yes (open weights) | $0 (self-hosted) | Compute cost only | Self-hosted via GitHub/Hugging Face |
The Google AI Ultra pricing is worth dwelling on because it changed recently. Google originally launched Ultra at $249.99 a month, then restructured the lineup into a $100-a-month tier with limited premium features, and a $200-a-month tier that includes Project Genie access alongside Gemini 3.1 Pro and other top-shelf tools, according to Engadget’s coverage of the price restructuring. If you only want Genie 3, the $100 tier won’t get you there — you need the full $200 plan.
Marble’s structure looks more like a typical SaaS product: a genuine free tier for experimentation, then $20, $35, and $95 monthly tiers that unlock higher resolution, commercial usage rights (starting at Pro), and API credits for teams that want to generate worlds programmatically rather than through the web interface. NVIDIA and Tencent both give away their model weights outright under open licenses, so the real cost is whatever GPU time you spend running inference, which for anything world-model-scale is not trivial — expect to need a workstation-class or datacenter GPU to get usable throughput.
Google DeepMind Genie 3: The Interactive World Simulator
Genie 3 is DeepMind’s bet that the fastest way to build general-purpose embodied AI is to first teach a model to generate the environments that AI will act inside. Access started rolling out on January 29, 2026, exclusively to US-based Google AI Ultra subscribers aged 18 and older — a narrow, controlled release consistent with how Google handled earlier Genie research previews.
Reported specs put Genie 3’s output at roughly 720p resolution and 24 frames per second, with about 60 seconds of interactive memory before the simulation’s coherence starts to degrade. That one-minute window is the single biggest technical constraint on the product right now: you can walk into a Genie 3 world, look around, and interact with it in real time, but you can’t yet treat it as a persistent, savable environment the way you could a Marble scene.
What Genie 3 does well is real-time playability. DeepMind’s own framing — “experimenting with infinite, interactive worlds” — captures the pitch: this isn’t a tool for exporting 3D assets, it’s a tool for stepping inside a generated place and moving around in it the way you would in a video game, except the game world didn’t exist five seconds before you typed the prompt. That makes it the most consumer-facing, demo-friendly product in this comparison, even though the $200-a-month price and US-only, 18-plus gating keep the actual user base small for now.
World Labs Marble: The Commercial 3D World Builder
World Labs, co-founded by Stanford computer vision pioneer Fei-Fei Li, launched Marble in limited beta in November 2025 and moved to a formal commercial launch in February 2026. By April, the company had shipped Marble 1.1 and a Plus variant, adding refinements to output quality and export options.
Marble takes a different technical approach than Genie 3: instead of a live, session-based interactive stream, it generates persistent 3D scenes from text, images, video, or 360-degree panoramas, then lets you export them as Gaussian splats — a rendering format increasingly popular in VFX and game-engine pipelines because it captures photorealistic detail without the manual modeling work of traditional 3D assets. That export capability is the feature that separates Marble from a pure demo: a studio can generate an environment in Marble, pull the splat file, and drop it into a game engine or a custom renderer for further production work.
Pricing follows a conventional freemium ladder: free tier, $20 Standard, $35 Pro, and $95 Max per month, with commercial usage rights kicking in at the Pro tier. There’s also a developer-facing API priced at $1 per 1,250 credits, which lets teams generate worlds programmatically rather than one at a time through the web app. World Labs has raised roughly $1.23 billion to build this out, a scale of funding that puts it closer to a foundation-model lab than a typical creative-tools startup, and reflects investor conviction that persistent 3D world generation has a real production market in gaming, VFX, and simulation.
NVIDIA Cosmos: World Models for Robots and Self-Driving Cars
NVIDIA took a completely different path with Cosmos. Announced at CES in January 2025 and expanded through a major release in March of that year, Cosmos isn’t trying to be a consumer product at all — it’s infrastructure. The platform ships under an open model license, with weights and a fine-tuning framework downloadable from Hugging Face and NVIDIA’s own NGC catalog, plus preview access through NVIDIA’s API catalog and Google Cloud’s Vertex AI Model Garden.
The Cosmos family breaks into named components: Cosmos Predict and Cosmos Transfer shipped openly, while Cosmos Reason arrived in early access. By 2026, industry roundups describe Cosmos 3 as the flagship, using what NVIDIA calls a mixture-of-transformers architecture that unifies vision reasoning, world generation, and action prediction into a single system. You can find the full technical rundown on NVIDIA’s Cosmos developer page.
The entire pitch is physics accuracy over visual polish. Cosmos generates physics-based synthetic video from text, images, existing video, or actual robot sensor and motion data, and NVIDIA integrates it with Omniverse blueprints to build synthetic data pipelines for post-training robots and autonomous vehicles. The point isn’t to make something that looks good in a demo reel — it’s to generate the rare, dangerous, or expensive-to-film scenarios (a child running into a street, a robot arm slipping on a wet floor) that a real camera crew could never safely or affordably capture at scale.
Decart Oasis and Lucy: Real-Time Video Without the Wait
Decart built its reputation on speed. The company’s original Oasis model, released in November 2024, was marketed as the first real-time video model and reportedly drew more than a million users within three days of launch — an early signal that “generate a playable world instantly” had real consumer pull, even before the “world model” label became an industry buzzword.
By 2026, Decart had split its product line in two. Lucy targets real-time immersive experiences for gaming, e-commerce, and advertising, with sub-30-millisecond response times — fast enough to feel instantaneous during interactive use. Oasis pivoted toward physical AI, with the June 2026 Oasis 3 release generating photorealistic multi-camera driving environments sold via API specifically to autonomous-vehicle developers who need to simulate rare road scenarios at scale, the same long-tail problem NVIDIA Cosmos targets from a different angle.
Decart’s funding trajectory reflects how seriously the automotive and chip industries are taking this space: a $100 million Series B in August 2025 valued the company at $3.1 billion, led by Sequoia Capital with Benchmark, Zeev Ventures, and Aleph VC participating. Nine months later, a $300 million round in May 2026 — led by Radical Ventures with NVIDIA, Adobe Ventures, Toyota Ventures, and eBay Ventures joining — pushed the valuation to roughly $4 billion. Total funding across five disclosed rounds now sits above $450 million. You can read more about the company’s direction on Decart’s official site.
Tencent HunyuanWorld: The Open-Source Path
Tencent has been the most prolific shipper in this category, and the only major lab treating world models as a genuinely open-source project rather than a gated product. HunyuanWorld-1.0 launched in mid-2025, built around panoramic proxy generation, semantic layering, and hierarchical 3D reconstruction to produce 360-degree, scene-scale environments from text or images. Component models like HunyuanWorld-PanoDiT-Text are listed publicly on GitHub, with the text-to-panorama model alone weighing in around 478 MB — small enough to run on consumer hardware, unlike the multi-billion-parameter LLMs dominating other AI headlines this year.
The lineage moved fast. HunyuanWorld-1.1, nicknamed WorldMirror, added support for building 3D worlds from video or multi-view images instead of just text and single images. HunyuanWorld-1.5, released December 18, 2025 and branded WorldPlay, shifted focus to real-time interactive world creation — Tencent’s answer to Genie 3’s live-simulation pitch. A FlashWorld variant compressed single-GPU 3D Gaussian Splat generation down to five to ten seconds. Then came HY-World-2.0 on April 16, 2026, described in Tencent’s own materials as a state-of-the-art multi-modal 3D world framework that handles both generation and reconstruction, outputting either mesh or Gaussian splat formats. The full project is browsable on Tencent’s HunyuanWorld GitHub repository.
Because it’s open-weight and self-hostable, HY-World is the model most likely to show up inside smaller studios, academic labs, and indie VR/AR projects that can’t justify a $95-a-month Marble subscription or don’t want to build against a closed API. The tradeoff is support: there’s no vendor on the other end of a support ticket, and you’re responsible for your own GPU infrastructure.
Benchmarks and Demo Comparisons: What the Data Actually Shows
This is the category’s biggest limitation for anyone doing due diligence: there is no standardized, third-party benchmark leaderboard for world models the way there is for LLMs on SWE-bench or MATH-500. Every comparison point available right now is demo-based or self-reported by the vendor, not validated by an independent evaluation body.
What we can compare directly, using each company’s own published figures: Genie 3 reports roughly 720p output at 24 fps with about 60 seconds of interactive coherence before drift sets in. Decart’s Lucy claims sub-30-millisecond response latency, positioning it as the fastest real-time option in this set, though Decart hasn’t published a resolution or fps figure to compare against Genie 3’s numbers directly. Tencent’s FlashWorld variant claims 5-to-10-second single-GPU generation for a full 3D Gaussian Splat scene, which is a throughput claim rather than a quality claim — it says nothing about how that scene looks compared to a Marble or Genie 3 output. NVIDIA and World Labs have not published resolution, fps, or latency numbers at all, framing their products instead around capability (physics accuracy, export formats, commercial usage rights) rather than raw performance metrics.
The practical takeaway: don’t trust any headline claiming one world model is objectively “faster” or “better” than another right now. Until an independent group publishes a standardized world-model benchmark — something several AI research labs have signaled interest in building — the honest comparison is on access model, output format, and fit for your specific use case, not raw performance.
Five Real-World Use Cases for AI World Models
World models aren’t a solution looking for a problem — several industries already have concrete, funded use cases in production or pilot testing.
- Autonomous vehicle simulation. NVIDIA Cosmos and Decart’s Oasis 3 both generate photorealistic, multi-camera driving environments specifically so AV developers can simulate rare, dangerous scenarios — a jaywalker at night, black ice on a highway curve — without needing a real car and a real risk to capture the footage.
- Robotics training data. NVIDIA pairs Cosmos with Omniverse blueprints to generate physics-accurate synthetic data for training robot perception and manipulation models, reducing how much real-world trial-and-error a robotics team needs before deployment.
- Game and VFX previsualization. Marble’s Gaussian splat export makes it usable inside existing production pipelines — a studio can generate an environment concept in minutes instead of having a 3D artist block it out from scratch, then refine from there.
- Interactive advertising and e-commerce. Decart’s Lucy targets sub-30ms real-time generation for gaming, advertising, and e-commerce experiences where a shopper might want to generate and explore a custom room layout or product environment on the fly.
- Academic and indie VR/AR research. Tencent’s fully open-source HunyuanWorld line gives university labs and solo developers a self-hostable option, avoiding both the $200-a-month Genie 3 paywall and the closed-API model that Decart and NVIDIA’s hosted services use.
Migration Guide: Moving From Traditional 3D Pipelines to World Models
If your team currently builds environments by hand in a traditional 3D pipeline and wants to evaluate world models as a supplement or replacement for part of that workflow, here’s a practical path.
- Define the deliverable format first. If your pipeline needs a mesh or Gaussian splat you can import into a game engine, Marble or Tencent’s HY-World are your realistic starting points — Genie 3 currently doesn’t offer a comparable export path since it’s built around live interactive sessions, not saved assets.
- Start on the free tier before committing budget. Marble’s free tier and NVIDIA Cosmos’s open-license download both let you test output quality against your actual use case before you pay for Marble Pro commercial rights or provision GPU infrastructure for self-hosted Cosmos.
- Benchmark against your own content, not vendor demos. Since there’s no independent leaderboard for this category yet, generate the same environment brief across two or three models and compare the outputs side by side on your own criteria — resolution, geometric consistency, export fidelity.
- Check commercial usage rights before shipping anything. Marble’s free and Standard tiers don’t include commercial usage rights — that kicks in only at the $35/mo Pro tier and above. Confirm licensing terms for whichever model you pick before putting generated assets into a shipped product.
- Budget for compute if you go open-weight. NVIDIA Cosmos and Tencent HY-World are free to download, but neither is free to run at production scale — plan for workstation-class or datacenter GPU costs, which can exceed a Marble or Genie 3 subscription depending on your generation volume.
- Treat robotics/AV use cases as a separate track. If your goal is synthetic training data for physical AI rather than creative or entertainment content, NVIDIA Cosmos and Decart’s Oasis 3 are purpose-built for that; Marble and Genie 3 are not optimized for the sensor-data and physics-accuracy requirements robotics teams need.
Pros and Cons: Genie 3
Pros: Real-time interactivity that feels like stepping into a generated video game world; backed by Google DeepMind’s research and compute resources; bundled with the rest of the Google AI Ultra suite, including Gemini 3.1 Pro access.
Cons: $200-a-month price is the steepest entry point in this comparison; US-only and 18-plus at launch, cutting off most of the global audience; 60-second memory window limits it to short, disposable sessions rather than persistent, saveable worlds; no public API for developers.
Pros and Cons: World Labs Marble
Pros: Genuine free tier lowers the barrier to entry; Gaussian splat export slots into existing VFX and game-engine pipelines; clear, published pricing ladder; API access for programmatic generation.
Cons: Commercial usage rights are locked behind the $35/mo Pro tier, so hobbyist output can’t be monetized without upgrading; not built for real-time interactivity the way Genie 3 or Decart’s Lucy are; a relatively young company (roughly two years old) building on $1.23 billion in investor expectations it still has to prove out commercially.
Pros and Cons: NVIDIA Cosmos
Pros: Open model license means no subscription cost and full control over deployment; purpose-built for robotics and AV synthetic data, a use case none of the other four models target as directly; backed by NVIDIA’s Omniverse ecosystem and NIM microservice deployment path.
Cons: No consumer-facing product — this is infrastructure for engineering teams, not a tool you open and start creating in; requires your own GPU compute, and NVIDIA hasn’t published clear pricing for hosted/NIM deployment; resolution, fps, and other performance specs aren’t publicly disclosed, making it hard to compare output quality directly against Marble or Genie 3.
Pros and Cons: Decart Oasis / Lucy
Pros: Sub-30ms response time on Lucy is the fastest real-time claim in this comparison; Oasis 3’s multi-camera driving simulation is a strong, focused fit for AV developers; heavyweight strategic investors (NVIDIA, Toyota Ventures, Adobe Ventures) signal serious industry confidence.
Cons: Fully closed and commercial, with no public flat-rate pricing published; credit-based API billing makes cost forecasting harder than a flat monthly subscription; smallest public technical documentation footprint of the five models covered here, making independent evaluation difficult before you commit to a contract.
Pros and Cons: Tencent HunyuanWorld / HY-World 2.0
Pros: Fully open-source with weights on GitHub and Hugging Face; free to download and self-host, no subscription or credit system; fastest generation claim in the category via the FlashWorld variant’s 5-to-10-second single-GPU output; broad input support spanning text, single-view images, multi-view images, and video.
Cons: No commercial support contract or vendor accountability if something breaks in production; self-hosting means you own the GPU infrastructure cost and complexity; English-language documentation and community support lag behind the primarily Chinese-language developer base.
The Verdict: Which AI World Model Should You Use?
There isn’t a single winner here, because these five products aren’t really competing for the same job. If you want the closest thing to walking around inside an AI-generated video game right now and don’t mind paying for the privilege, Genie 3 is the most polished real-time experience — but the $200-a-month price tag and 60-second memory ceiling make it a demo tool more than a production one today.
If you’re a studio or solo creator who needs an actual exportable 3D asset for a game, VFX shot, or VR experience, Marble is the most production-ready option on the market. The free tier lets you test before you pay, and the Gaussian splat export means the output isn’t a dead end — it plugs into pipelines your team probably already uses.
If your use case is robotics or autonomous vehicles, skip the consumer-facing tools entirely. NVIDIA Cosmos and Decart’s Oasis 3 are both purpose-built for exactly that job, generating physics-accurate synthetic data and multi-camera driving environments respectively, and NVIDIA’s open licensing means you can start testing without negotiating a contract first.
And if budget is the deciding factor or you want full control over your infrastructure, Tencent’s HunyuanWorld line is the only fully open-source, self-hostable option in this comparison, with a release cadence — four major versions in under a year — that suggests Tencent is treating this as a genuine long-term platform bet, not a side project.
The bigger story is what this category signals about where AI investment is heading next. Genie 3, Marble, Cosmos, Oasis, and HY-World collectively represent well over $2 billion in disclosed funding and internal investment from companies with no shortage of capital to deploy elsewhere. That’s a strong bet that the next major leap in AI capability won’t come from a bigger chatbot — it’ll come from AI systems that understand and can generate physical space, the same skill every robot, self-driving car, and game engine ultimately needs.
Frequently Asked Questions
What is an AI world model?
An AI world model is a generative system that simulates a spatially coherent 3D or interactive video environment, rather than generating text, a static image, or a short video clip. Unlike a chatbot such as GPT-5.6 or Gemini, which reasons over language, a world model reasons over physical space — letting you walk through, drive in, or extract assets from a generated place.
Is Google Genie 3 available to the public?
Yes, but with restrictions. As of September 2026, Genie 3 access requires a Google AI Ultra subscription at the $200-a-month tier, and is limited to US-based users aged 18 and older. Google’s lower $100-a-month Ultra tier does not include Genie 3 access.
How much does World Labs Marble cost?
Marble offers a free tier for basic use, then paid plans at $20/month (Standard), $35/month (Pro), and $95/month (Max). Commercial usage rights for generated content begin at the Pro tier. An API is also available at $1 per 1,250 credits for developers who want to generate worlds programmatically.
Is NVIDIA Cosmos free to use?
The Cosmos model weights are free to download under NVIDIA’s open model license via Hugging Face and NVIDIA’s NGC catalog. You’ll still need to pay for the GPU compute required to run inference, whether that’s your own hardware or a cloud provider.
Which world model is best for autonomous vehicle simulation?
NVIDIA Cosmos and Decart’s Oasis 3 are both purpose-built for this use case, generating physics-accurate, multi-camera driving environments that AV developers use to simulate rare or dangerous road scenarios without real-world risk. Cosmos has the advantage of open licensing, while Oasis 3 is sold as a closed commercial API.
Can I use AI world models to generate assets for a game engine?
Yes. World Labs’ Marble and Tencent’s HunyuanWorld models both export as 3D Gaussian splats or mesh formats that can be imported into common game engines. Genie 3, by contrast, is built around live interactive sessions and does not currently offer a comparable persistent export path.
Are there independent benchmarks comparing these world models?
Not yet. As of September 2026, there is no standardized, third-party benchmark leaderboard for world models comparable to SWE-bench for coding models or MATH-500 for reasoning models. All performance figures currently available are self-reported by each vendor, so treat cross-model performance claims with appropriate caution until independent evaluation frameworks emerge.
Who is Yann LeCun’s connection to the world model category?
Yann LeCun left Meta to found a new venture centered on world models, arguing they represent a more promising path toward general-purpose AI than scaling large language models further. The startup has raised over $1 billion, underscoring how seriously investors are taking the world-model thesis as a potential successor architecture to today’s LLMs.
Related Coverage
- Tencent Hy3 vs GLM-5.3-Flash vs Kimi K3: 20x Price Gap [2026]
- Chatbase vs Voiceflow vs Botpress: $350 Pricing Gap [2026]
- Meta Hatch AI Agents, Watermelon Model at $199.99 [2026]
- Gemma 4 vs Phi-4 Mini vs Qwen3.5: 4.75x Size Gap [2026]
- How to Use Qwen3.8 Flash API: 1M Context, 12 Steps [2026]
