AI startup Descart on Wednesday announced Oasis 3, its latest interactive world model that can generate photorealistic driving environments in real time, TechCrunch has learned exclusively. This model is currently available via API.
The startup is initially targeting self-driving car companies that need to simulate rare driving scenarios at scale, and plans to expand into robotics and other physical AI applications. But the bigger bet is on the developers. Decart is looking to build a developer ecosystem around world models, similar to what OpenAI did with language models, by providing API access from day one.
“This will be the first usable world model that people can actually program on,” Decart co-founder and CEO Dean Leitersdorf told TechCrunch. “I think a whole developer community will emerge on top of this.”
The startup already has a community of over 100,000 developers, many of whom are building products based on the real-time video model Lucy, primarily in e-commerce and live streaming. Oasis 3 is based on that foundational model and represents the company’s commitment to physical AI. Decart said access will be priced at $0.02 per second, with enterprise pricing determined by use case.
Decart is playing in an increasingly crowded world model arena. Last year, Google released Genie 3 in research preview, Fei-Fei Li’s World Labs launched Marble for commercial use cases, and video generation startups like Luma and Runway are also converting physics-aware video models into world models.
The release of Oasis 3 comes weeks after two-year-old Descart raised $300 million, which Leitersdorf said followed a “significant increase in demand for the models we built” in e-commerce, live streaming, and physical AI. The round raised Decart’s valuation to nearly $4 billion, with participation from strategic investors including Toyota, Adobe, and eBay. All of these companies are potential customers, Reitersdorf said. Existing investor NVIDIA also participated in the round.
Oasis 3’s advantage lies in its model photorealism and infinite generation capabilities. This is due to efficiency wizardry on the Descart side, leveraging another of Descart’s flagship products, the DOS (Decart Optimization Stack) software. This allows us to run our models efficiently on hardware from Nvidia, Amazon, and Google, making them significantly cheaper to run than our competitors.
“It’s built on top of our entire real-time stack and optimized down to the hardware,” Leitersdorf said. “Because we are highly vertically integrated, we can run these models more than an order of magnitude cheaper than other companies in the industry.”
Reitersdorf says the startup’s model is so efficient that it has spent “significantly” less than $100 million over its lifetime.
Oasis 3 produces a physically accurate multi-camera environment (one front and two sides) for training and testing systems. And instead of offering limited demos and research previews, Descart allows developers to generate an infinite number of scenarios, making it perfect for self-driving car developers looking to try out as many edge cases as possible.
Compared to other models I’ve tried, including Google’s Genie 3 and World Labs’ Marble, the Oasis 3 provides the most photorealistic environment from a single text prompt I’ve ever seen. And the fact that you can interact with them for hours suggests a level of efficiency that Descart’s rivals may lack.
However, if you let the world generate for too long, the model will also degrade significantly.
In my testing, I found that the system was able to consistently set up strong initial scenes that matched the prompts, but the integrity of the theme quickly deteriorated as I moved through the world. In the morning, I prompted it to generate a New York City street, and it generated it beautifully. However, as I drove, the environment looked more like a standard version of a western city than New York.
When I turned around and tried to go back to the first intersection, it disappeared and was replaced by a completely new environment. On top of that, the controls weren’t very responsive and I often lost control of where the car was moving (this was also a common shortcoming with other Worldwide models I tested). The experience felt more like a dream-like, disjointed stream of consciousness than a coherent simulation, and quickly became meaningless.
Another problem I’ve seen with other world models is that cars just pass between other cars and the model doesn’t properly simulate the physics in the environment. Professor Reitersdorf calls this “a major research problem that we are currently solving” and attributes it to the fact that “there is significantly more data on good driving compared to accidents.”
One of the things that makes this physics so hard to be consistent is the fundamentals of how this model of the world works. Oasis 3 is autoregressive. That is, it generates one frame at a time and looks back at what was previously generated to determine what comes next. This is an important architectural feature of many world models, and is also computationally intensive.
To maintain consistency, Leitersdorf said the Descart team is working to improve the memory length of the model.
“Each frame we generate is approximately 8,000 tokens,” he said. “If you generate this at tens of frames per second, you’re generating hundreds of thousands of tokens per second. The context window fills up quickly. We’re investigating how to run longer contexts and store millions more tokens, and how to compress memory into fewer tokens.”
Leitersdorf believes that the next version of the model may partially resolve the consistency issue, allowing users to generate worlds based on videos of the environment rather than images. He acknowledged that global models are still in their infancy as a field.
Still, the founders are less focused on the current limitations of their technology and more focused on what happens when developers get their hands on it.
“We go back to the early days of LLM when OpenAI invented APIs for models,” he said, pointing to the emergence of a developer community that advanced the field by finding and building new use cases.
“When we talk again in three months, we’ll say, ‘We’ve got 100 developers here, and they’ve all built 100 different applications on Oasis, surprising us all,'” he said.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
