Decart Unveils Oasis 3 World Model for Photorealistic Driving Simulation

Decart, an AI startup known for real‑time generative video technology, this week announced Oasis 3, its next‑generation world model capable of generating hours of photorealistic driving environments for use in autonomous vehicle testing, robotics research and other physical AI applications. Available today via API, the model represents a step forward in interactive simulation, but comes with important limitations that industry players and developers should weigh.

Oasis 3 builds on Decart’s broader world‑modeling efforts, which previously enabled real‑time video generation and interactive environments, applying those capabilities to large‑scale driving scenes that include dynamic weather, traffic and environmental variation. Its release follows a $300 million funding round that lifted Decart’s valuation to nearly $4 billion and counts strategic investors such as Toyota, Adobe and Nvidia.

Photorealism Meets Real‑World Simulation

Unlike static 3D maps or handcrafted autonomous test tracks, Oasis 3’s simulated worlds are generated on the fly and can run for extended periods, producing multi‑camera, video‑like outputs that resemble real driving scenarios. The model supports continuous simulation from simple text prompts, enabling developers to produce hours of photorealistic road footage for training perception systems and edge‑case testing without the need for traditional synthetic environment setups.

This capability places Oasis 3 in the same emerging category as other world models, AI systems designed to simulate rich environments for planning, testing and agent learning, but with a specific focus on automotive and physical AI use cases. Google’s Genie 3, for example, similarly creates immersive environments from prompts and is being applied to robotics and educational tools, while Waymo’s custom world model uses a Genie‑derived core to generate rare driving scenarios for safety validation.

Practical Limits and Caveats

Despite its promising photorealism and duration, Oasis 3 has notable caveats. Extended simulations tend to degrade in consistency over time, especially in long continuous runs, a common challenge in current generation world models, which can introduce artifacts or continuity errors that make some sequences less useful for rigorous testing. Heavy computational demands also mean that powerful hardware and substantial cloud resources are often required to generate these environments at scale, limiting accessibility for smaller teams and individual developers.

These limitations echo broader industry observations about world models: while they can create visually impressive, immersive scenarios, ensuring physical accuracy, causal consistency and long‑term stability remains a work in progress. Models like Google’s Genie and others are still advancing toward reliably embedding real‑world physics and environmental laws into their simulations, something critical for high‑stakes applications like autonomous driving testing.

Developer Ecosystem Focus

Decart’s broader strategy with Oasis 3 is not just to serve autonomous vehicle labs or research teams, but to seed a developer ecosystem around world models. The company has more than 100,000 developers already building on its existing real‑time video models, particularly in e‑commerce and livestreaming, and hopes that offering API access from launch will foster novel applications across sectors.

Industry experts say that world models like Oasis 3 and others, including Google’s Genie‑powered systems and startup offerings, could become foundational tools for simulation‑based machine learning, safety validation and robotics development if challenges around accuracy and efficiency are solved. Doing so would help bridge a critical gap between synthetic training and real‑world deployment, especially in fields where gathering rare or dangerous data, like severe weather driving, is impractical or unsafe.