AI startup Decart has recently introduced Oasis 3, its advanced interactive world model capable of generating photorealistic driving environments in real time, as exclusively reported by TechCrunch. This innovative model is currently accessible through an API.
The company's initial strategy targets autonomous vehicle manufacturers requiring extensive simulation of rare driving scenarios, with future plans to expand into robotics and other physical AI applications. However, Decart's primary focus is on fostering a robust developer ecosystem around world models, mirroring OpenAI's successful approach with language models, by offering API access from its inception.
Dean Leitersdorf, co-founder and CEO of Decart, shared his vision with TechCrunch, stating, “It’s going to be the first usable world model that people can actually program on top of.” He added, “I think there’s going to be an entire developer community that emerges on top of this.”
Decart already boasts a vibrant community of over 100,000 developers, many of whom are actively building products using its real-time video model, Lucy, predominantly in e-commerce and live streaming. Oasis 3 builds upon this foundational model, marking the company’s strategic push into the realm of physical AI. Access to Oasis 3 is priced at $0.02 per second, with enterprise rates customized based on specific use cases, according to Decart.
Decart enters a competitive landscape within the world model sector. Last year saw Google's release of Genie 3 in a research preview, while Fei-Fei Li’s World Labs launched Marble for commercial applications. Additionally, video generation companies such as Luma and Runway are actively transforming their physics-aware video models into comprehensive world models.
The launch of Oasis 3 follows Decart's recent announcement of a $300 million funding round, secured just weeks prior. Leitersdorf indicated that this investment was driven by “huge demand increases for the models we built” across e-commerce, live streaming, and physical AI. This round propelled Decart's valuation to nearly $4 billion and attracted strategic investors including Toyota, Adobe, and eBay, all identified by Leitersdorf as potential customers. Nvidia, an existing investor, also participated in the funding.
Oasis 3 distinguishes itself through the exceptional photorealism of its models and its capacity for infinite generation. This is attributed to Decart’s proprietary efficiency innovations, particularly its DOS (Decart Optimization Stack) software. DOS enables models to operate efficiently on Nvidia, Amazon, and Google hardware, significantly reducing operational costs compared to competitors.
“This is built on top of our entire real-time stack, which we optimize all the way down to the hardware,” Leitersdorf explained. He emphasized, “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.”
According to Leitersdorf, the remarkable efficiency of Decart’s models has resulted in the company expending "drastically less" than $100 million throughout its operational history.
Oasis 3 is designed to generate physically accurate, multi-camera environments—featuring one front-facing and two side-facing views—ideal for training and testing AI systems. Unlike offerings with limited demos or research previews, Decart empowers developers with the ability to generate scenarios infinitely.
In comparative testing, Oasis 3 stood out by producing the most photorealistic environments from a single text prompt, surpassing models like Google’s Genie 3 or World Labs’s Marble. The ability to interact with these environments for extended periods also points to an efficiency level that competitors may currently lack.
However, prolonged generation of a world does lead to a noticeable degradation in the model’s quality.
During my evaluation, the system consistently established a strong initial scene that accurately matched the prompt. Yet, the thematic integrity rapidly diminished as I navigated through the generated world. For instance, when prompted to create a New York City street in the morning, the initial rendition was beautiful. However, as I drove along, the environment quickly lost its distinct New York characteristics, becoming more akin to a generic urban setting in any Western city.
Attempts to turn around and return to the starting intersection proved futile, as it had been replaced by an entirely new environment. Furthermore, the controls exhibited poor responsiveness, frequently leading to a loss of command over the vehicle's movement—a common issue observed in other world models I've tested. The overall experience felt less like a cohesive simulation and more like a dream-like, disjointed stream of consciousness that quickly became nonsensical.
Another persistent issue, also noted in other world models, is the car’s tendency to pass through other vehicles, indicating a failure in proper physics simulation within the environment. Leitersdorf acknowledged this as a “major research problem that we’re cracking now,” attributing it to a significant imbalance in available data, where "there’s drastically more data on good driving compared to accidents.”
The challenge of maintaining physics consistency is inherent to Oasis 3’s design. As an auto-regressive model, it generates one frame at a time, referencing previously generated frames to determine subsequent outputs. This architectural characteristic is common among many world models and is notably compute-intensive.
To address consistency, Leitersdorf stated that the Decart team is focused on enhancing the model's memory capacity.
“Every frame we generate is roughly 8,000 tokens,” he detailed. “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 anticipates that the consistency issue may be partially resolved in the model’s upcoming version, which will allow users to initiate world generation from a video of an environment rather than a static image. He candidly admitted that the field of world models is still in its nascent stages.
Despite current limitations, the founder remains more focused on the transformative potential once developers begin interacting with the technology.
“It takes me back to the early days of LLMs, when OpenAI invented the API for models,” he remarked, drawing parallels to how a burgeoning developer community propelled the field forward by discovering and implementing novel use cases.
He confidently predicted, “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.’”
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.