“It’s going to be the first usable world model that people can actually program on top of,” Decart co-founder and CEO Dean Leitersdorf told TechCrunch. “I think there’s going to be an entire developer community that emerges on top of this.”
Oasis 3 builds on Decart’s existing real-time video model, Lucy, which the company says is already used by more than 100,000 developers across areas such as e-commerce and live streaming. Access to the new model starts at $0.02 per second, with enterprise pricing based on individual use cases.
The release comes shortly after Decart raised $300 million at a valuation approaching $4 billion. The funding round included investors such as Toyota, Adobe, eBay, and Nvidia, which had previously backed the company.
Decart enters a growing market for world models, joining competitors including Google’s Genie 3, World Labs’ Marble, and technologies being developed by video generation companies such as Luma and Runway.
According to the company, Oasis 3’s main advantage is its ability to generate photorealistic environments indefinitely while maintaining low operating costs. Leitersdorf attributed that efficiency to the Decart Optimization Stack, or DOS, software platform, which is designed to optimize model performance across Nvidia, Amazon, and Google infrastructure.
“This is built on top of our entire real-time stack, which we optimize all the way down to the hardware,” Leitersdorf said. “By being so vertically integrated, we’re able to be more than an order of magnitude cheaper than anyone else in the industry in order to run these models.”
The model generates three-camera driving environments consisting of a front-facing view and two side views, allowing developers to test systems across a range of road conditions and scenarios. Unlike many research previews, Decart is making the system broadly available through an API from launch.
Early testing highlighted both the strengths and limitations of the technology. Oasis 3 was able to create highly realistic environments from simple prompts and maintain interactive simulations for extended periods. However, scene consistency deteriorated over time, locations changed unexpectedly, and the system sometimes failed to model physical interactions accurately, including collisions with other vehicles.
Leitersdorf acknowledged that maintaining long-term consistency remains one of the industry’s biggest technical challenges. Because Oasis 3 generates environments frame by frame, memory limitations become increasingly important as simulations grow longer.
“Every frame we generate is roughly 8,000 tokens,” he said. “Generating this at tens of frames per second — that’s hundreds of thousands of tokens per second. The context window fills up very quickly.”
Decart says future versions of Oasis will focus on improving memory, consistency, and environmental persistence. The company is also working toward allowing developers to generate interactive worlds directly from video inputs rather than static images.
Despite the technology’s current limitations, Decart sees API access as the key differentiator. The company believes developers will discover new applications for world models in the same way developers helped expand the use cases for large language models after API access became widely available.
This analysis is based on reporting from TechCrunch.
Images courtesy of Decart.
This article was generated with AI assistance and reviewed for accuracy and quality.
