Decart’s world model simulates hours of photorealistic driving, with caveats

AI startup Decart has officially launched Oasis 3, an interactive world model capable of generating photorealistic driving environments in real time, as exclusively reported by TechCrunch. The model is now accessible through an API, marking a significant step for the company.

The startup is initially targeting autonomous vehicle companies that require large-scale simulation of rare driving scenarios, with plans to expand into robotics and other physical AI applications. However, the larger ambition revolves around developers. By offering API access from the outset, Decart aims to cultivate a developer ecosystem around world models, similar to how OpenAI built a community around language models.

“It’s going to be the first usable world model that people can actually program on top of,” said Dean Leitersdorf, co-founder and CEO of Decart. “I think there’s going to be an entire developer community that emerges on top of this.”

Decart already boasts a community of over 100,000 developers, many of whom are building products on its real-time video model, Lucy, primarily in e-commerce and live streaming. Oasis 3 is built on that foundation model and represents the company’s push into physical AI. Access is priced at $0.02 per second, with enterprise pricing varying by use case.

The company enters a crowded field. Last year, Google released Genie 3 in research preview, Fei-Fei Li’s World Labs launched Marble for commercial use, and video generation startups like Luma and Runway are also translating their physics-aware video models into world models.

Oasis 3’s release follows Decart’s $300 million funding round a few weeks ago, which Leitersdorf says came after “huge demand increases for the models we built” in e-commerce, live streaming, and physical AI. The round valued the two-year-old startup at nearly $4 billion and brought in strategic investors including Toyota, Adobe, and eBay, all potential customers. Nvidia, an existing investor, also participated.

The model’s edge lies in its photo-realism and infinite generation capability, achieved through efficiency optimizations powered by Decart’s other main product: the DOS (Decart Optimization Stack) software. This stack allows models to run efficiently on Nvidia, Amazon, and Google hardware, making them far less expensive to operate than competitors.

“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 startup’s models are so efficient, per Leitersdorf, that it has burned through “drastically less” than $100 million in its lifetime.

Oasis 3 generates physically accurate, multi-camera environments , one front-facing and two side-facing , for training and testing systems. Instead of offering limited demos or research previews, Decart allows developers to generate scenarios infinitely, ideal for autonomous vehicle developers seeking as many edge cases as possible.

Compared to other models I’ve tested, like Google’s Genie 3 or World Labs’s Marble, Oasis 3 delivers the most photorealistic environments from a single text prompt I’ve seen. The ability to interact with them for hours suggests a level of efficiency that rivals may lack.

However, the model degrades significantly over extended use. In my testing, it consistently set up a strong initial scene matching the prompt, but thematic integrity eroded rapidly as I moved through the world. A prompt for a New York City street in the morning produced a beautiful scene, but as I drove, the environment lost its distinct character, becoming a generic urban Western city. Attempting to return to the initial intersection failed, replaced by an entirely new environment. Controls felt unresponsive, and I often lost direction. The experience felt less like a coherent simulation and more like a dream-like, disjointed stream of consciousness.

Another issue, common among world models, is that the car drives through other vehicles, indicating the model doesn’t simulate physics properly. Leitersdorf calls this “a major research problem that we’re cracking now,” attributing it to the fact that “there’s drastically more data on good driving compared to accidents.”

The difficulty in maintaining physics consistency stems from Oasis 3’s auto-regressive architecture. It generates one frame at a time, looking back at previous frames to decide what comes next. This is a key feature of many world models and is compute-intensive.

To improve consistency, Leitersdorf says the team is working on extending the model’s memory.

“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. We’re researching how to do longer context to store millions more tokens, and how to compress the memory into fewer tokens.”

Leitersdorf believes the consistency issue may be partially solved in the next version, which will allow users to generate worlds based on a video of an environment rather than an image. He acknowledges that world models as a field are still early.

Still, the founder is more focused on the potential of developer adoption than current limitations.

“It takes me back to the early days of LLMs, when OpenAI invented the API 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 be like, ‘Here’s 100 developers that all built 100 different applications with Oasis that surprised all of us.’”