Decart’s new global model can simulate hours of real-world driving — with a few caveats

AI startup Decart on Wednesday unveiled Oasis 3, its latest interactive world model that can create realistic driving environments in real time, TechCrunch has learned exclusively. The form is currently available via the Application Programming Interface (API).

The startup is initially targeting autonomous vehicle companies that need to simulate rare driving scenarios at scale, and plans to expand into robotics and other physical AI applications. But the biggest bet is on developers: By providing API access from day one, Decart is trying to build a developer ecosystem around universal models much like what OpenAI did with language models.

“This will be the first universally usable model that people can actually program,” Dean Leitersdorf, co-founder and CEO of Decart, told TechCrunch. “I think there will be a whole developer community that will emerge on top of this.”

The startup already has a community of more than 100,000 developers, many of whom are building products on top of Lucy’s real-time video model, largely in e-commerce and live streaming. Oasis 3 builds on this basic model, and represents the company’s push into physical AI. Descartes said the access price is $0.02 per second, and enterprise pricing depends on use cases.

Descartes plays in an increasingly crowded global fashion show arena. Last year, Google released Genie 3 in search preview, Fei-Fei Li’s World Labs launched Marble for commercial use cases, and video production startups like Luma and Runway are also translating physics-aware video models into world models.

The release of Oasis 3 comes just weeks after two-year-old Decart raised $300 million, which Leitersdorf says came on the heels of “huge increases in demand for the models we’ve built” in e-commerce, live streaming and physical AI. The round boosted Descartes’ valuation to nearly $4 billion, and brought in a series of strategic investors such as Toyota, Adobe, and eBay. All of these companies are potential clients, Leitersdorf says. Nvidia, an existing investor, also participated in the round.

The advantage of Oasis 3 lies in the photo-realism of its models and its unlimited production capabilities. This is due to some efficiency skills on Decart’s part, supported by the company’s other flagship product: DOS software (Decart Optimization Stack) that allows models to run efficiently on Nvidia, Amazon and Google hardware, making their models much less expensive to run than competitors.

“This is built on top of our entire real-time stack, which we are optimizing down to the hardware,” Leitersdorf said. “By vertical integration, we are able to be much cheaper than anyone else in the industry to operate these models.”

The startup’s models are so effective, according to Leitersdorf, that it has spent “significantly less” than $100 million in its lifetime.

Oasis 3 creates physically accurate, multi-camera environments – one front and one side – for training and testing systems. Instead of offering limited demos and research previews, Decart allows developers to create infinite scenarios.

Compared to other models I’ve tried, like Google’s Genie 3 or World Labs’ Marble, the Oasis 3 delivers the most realistic environments through a single text message I’ve seen. And the fact that you can interact with them for hours indicates a level of efficiency that Decart’s competitors may lack.

But by letting you create a world for so long, the model also deteriorates significantly.

In my testing, I found that the system could always set up a strong initial scene that matched the prompt, but the thematic integrity quickly deteriorated as I moved through the world. I asked him to create a New York City street in the morning, and he did it beautifully. But as I drove, the environment looked less like New York and more like a standard version of any Western metropolitan city.

When I tried to turn around and return to the initial intersection, it disappeared and was replaced by a completely new environment. Furthermore, the controls aren’t responsive, and I often lost control of where the car was moving (again, a flaw shared by other global models I’ve tested). The experience seemed less like a coherent simulation and more like a disjointed, dream-like stream of consciousness that quickly became inconsequential.

Another problem, which I’ve also seen in other world models, is that the car will only pass other cars, which means the model doesn’t properly simulate the physics of the environment. Leitersdorf calls this “a major research problem that we are working on now,” and attributes this to the fact that “there is significantly more data on good driving than on accidents.”

Part of what makes this physical consistency difficult is that it is fundamental to how this world model works. Oasis 3 is auto-retrograde, meaning it generates one frame at a time, looking at what was generated previously to decide what comes next. This is a key architectural feature of many global models, and is compute-intensive as well.

To maintain consistency, Leitersdorf says Descartes’ team is working to improve the model’s memory length.

“Every frame we create is about 8,000 characters,” he said. “Building this at tens of frames per second — that’s hundreds of thousands of tokens per second. The context window fills up very quickly. We’re looking at how to make a longer context to store millions of tokens, and how to compress memory into a smaller number of tokens.”

Leitersdorf believes the consistency issue may be partially solved in the next version of the model, which will allow users to start creating worlds based on a video of an environment rather than an image. He acknowledged that global modeling as a field is still early.

However, the founder is less focused on the current limitations of his technology than what will happen when developers get their hands on it.

“It takes me back to the early days of the MBA, when OpenAI invented the modeling API,” he said, noting 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 be like: ‘Here are 100 developers who have all built 100 different apps with Oasis, which surprised us all,'” he said.