Decart launches Oasis 3 world model for real-time simulation

On Wednesday, AI startup Decart unveiled Oasis 3, an interactive world model—an AI model that simulates physical environments—that generates photorealistic driving environments in real time. The company is initially targeting autonomous vehicle companies for simulation scenarios, with plans to expand into robotics and other physical AI applications. The model, which is based on the Lucy real-time video model, is currently available to developers via an API priced at $0.02 per second.

By offering immediate API access, Decart aims to build a developer ecosystem around its technology. Dean Leitersdorf, co-founder and CEO of Decart, compared the approach to the early days of language models, stating, “I think there’s going to be an entire developer community that emerges on top of this.” This launch places Decart in competition with several technology companies and specialized startups developing physical AI models, including:

• Google, which previously released its Genie 3 model in research preview.
• World Labs, which launched its Marble model for commercial use cases.
• Video generation companies Luma and Runway, which are translating their models into world models.

The release follows a $300 million funding round that valued the two-year-old startup at nearly $4 billion. The round attracted strategic investors including Toyota, Adobe, and eBay, alongside existing investor Nvidia. Decart attributes its pricing to its Decart Optimization Stack (DOS) software, which optimizes model performance down to the hardware level. Leitersdorf noted that this vertical integration allows the company to run these models at a cost that is more than an order of magnitude cheaper than competitors. Consequently, the startup has burned through drastically less than $100 million in its lifetime.

While Oasis 3 generates physically accurate, multi-camera environments, it faces notable technical limitations. The model degrades significantly when generating worlds for long periods. For instance, in testing, a generated New York City street environment gradually lost its specific thematic integrity, reverting to a generic Western city before disappearing entirely. Additionally, the model does not simulate physics properly, causing cars to drive through other cars. Leitersdorf acknowledged this issue, calling it a “major research problem that we’re cracking now” due to a lack of accident data. Part of the difficulty stems from the model’s auto-regressive architecture—meaning it generates one frame at a time based on previous frames. With each frame requiring roughly 8,000 tokens, the context window fills up rapidly, prompting Decart to research memory compression and longer context windows.

Why it matters

Decart is attempting to build a developer ecosystem around world models by offering API access to its new Oasis 3 model, positioning itself to compete with major players like Google and World Labs in the physical AI space.