level: technical
ai startup decart launched oasis 3, an interactive world model that creates photorealistic driving environments from text prompts. the model runs in real time and is available through an api, priced at $0.02 per second. decart aims to serve autonomous vehicle companies needing to simulate rare scenarios at scale, with plans to expand into robotics and other physical ai applications. the company also hopes to build a developer community around world models, similar to how openai did with language models.
oasis 3 generates multi-camera views—front and two sides—and can run for hours due to efficiency gains from decart's optimization stack, which works across nvidia, amazon, and google hardware. this vertical integration makes the model cheaper to operate than competitors. however, testing revealed significant limitations: environments degrade over time, losing thematic consistency and becoming generic. the model lacks object awareness, allowing the car to drive through other vehicles, and controls are unresponsive, leading to a disjointed experience.
the model is auto-regressive, generating one frame at a time based on previous frames, which strains memory as context windows fill quickly. decart is researching longer context and memory compression to improve consistency. the next version may allow generation from video inputs. despite current flaws, decart sees oasis 3 as a foundation for developers to build applications, betting that an api-first approach will spur innovation in the early-stage world model field.
why it matters: world models like oasis 3 could enable scalable, low-cost simulation for training autonomous systems, but current consistency issues limit reliability for safety-critical applications.