AI company Decart revealed Oasis 3 on Wednesday, its newest interactive world model that produces photorealistic driving scenes in real time, TechCrunch has learned exclusively. The model is now offered through an API.

The company focuses first on autonomous vehicle firms needing to simulate uncommon driving situations at large scale, with plans to move into robotics and additional physical AI uses. Yet the main goal targets developers: By providing API access immediately, Decart aims to create a developer community around world models similar to OpenAI's approach with language models.

"This will serve as the first practical world model that users can truly build upon," Dean Leitersdorf, co-founder and CEO of Decart, told TechCrunch. "I expect a full developer community to form around this technology."

The startup already supports over 100,000 developers, many creating tools using its real-time video model Lucy, mainly for e-commerce and live streaming. Oasis 3 builds on that base model and marks the firm's move into physical AI. Pricing starts at $0.02 per second, while enterprise rates vary by application, Decart noted.

Decart enters a crowded field of world models. Last year Google launched Genie 3 in research preview, Fei-Fei Li's World Labs released Marble for commercial needs, and video firms such as Luma and Runway are adapting their physics-aware video models into world models.

Oasis 3 arrives weeks after two-year-old Decart secured $300 million, which Leitersdorf attributes to strong demand growth for its models in e-commerce, live streaming and physical AI. The funding raised the valuation to nearly $4 billion and added strategic backers including Toyota, Adobe and eBay. Leitersdorf views all as potential clients. Nvidia, already an investor, joined the round.

Oasis 3 stands out through photorealistic output and unlimited generation. This stems from efficiency advances via the company's DOS software, which runs models efficiently on Nvidia, Amazon and Google hardware, cutting costs far below rivals.

"We built this on our full real-time stack, optimized down to the hardware level," Leitersdorf said. "Vertical integration lets us run these models more than ten times cheaper than others in the space."

The models prove so efficient, Leitersdorf added, that lifetime spending has stayed well below $100 million.

Oasis 3 creates physically accurate multi-camera scenes - one front view and two side views - for training and testing. Rather than limited demos, Decart lets developers generate endless scenarios, suiting autonomous vehicle teams seeking many edge cases.

Against other models such as Google's Genie 3 or World Labs's Marble, Oasis 3 produces the most photorealistic scenes from a single text prompt observed so far. Extended interaction for hours points to efficiency advantages over competitors.

Yet extended generation leads to noticeable quality drop.

Loading the player...

Testing showed the system builds strong initial scenes matching the prompt, yet thematic consistency fades quickly during movement. A prompt for a New York City morning street produced a fine result, but further driving shifted the setting toward a generic urban Western cityscape.

Attempts to return to the starting intersection failed as it vanished into a fresh scene. Controls also respond slowly, often causing loss of direction, a common issue across tested world models. The result feels like a fragmented, dream-like flow that turns incoherent fast.

Another shared flaw is vehicles passing through others, showing weak physics simulation. Leitersdorf describes this as a key research focus, noting far more training data covers normal driving than collisions.

Physics consistency remains hard due to the model's auto-regressive design, which creates one frame at a time while referencing prior output. This approach, common in world models, demands heavy compute.

To boost consistency, Leitersdorf said the team works on extending memory length.

"Each frame uses about 8,000 tokens," he said. "At tens of frames per second this reaches hundreds of thousands of tokens quickly. The context fills fast. We study longer contexts to hold millions more tokens and ways to compress memory."

Leitersdorf expects partial fixes in the next version, which will support world generation from video input instead of images. He noted the field remains early stage.

The founder focuses more on future developer uses than present limits.

"This recalls early LLM days when OpenAI released its model API," he said, referencing how developers expanded the space with new applications.

"When we speak again in three months, expect 100 developers creating 100 surprising applications with Oasis," he said.