How Decart built Oasis 3, the first API-accessible world model for faster, more effective, and scalable robotic learning
The Tel Aviv-based frontier AI research lab is focused on driving the world’s most efficient AI, powering interactive experiences from live video to next-gen world models. Their work sits at the intersection of generative AI and physical AI, building the simulation infrastructure that autonomous systems and robotics teams depend on to develop and test in the real world.
The opportunity Decart is pursuing is massive: video powers the modern economy, yet pixel-based AI remains too slow, expensive, and limited to match the capabilities we’ve come to expect from text and code. Decart is closing that gap with world models that deliver sub-40 millisecond latency, run up to 1,000x faster than industry standards, and cost up to 400x less than competing approaches.
Oasis 3: Decart's flagship interactive world model for physical AI simulation and training
Oasis 3 generates realistic, controllable, multi-view simulation environments in real-time, where robots can learn, act, and improve across infinite scenarios. Give it a prompt and it returns an interactive environment you can navigate continuously while the world responds around you. As a world model, instead of rendering a scene from fixed rules, Oasis 3 generates realistic, dynamic visuals from learned representations of the physical world.
Oasis 3 is the API-accessible interactive world model built for Physical AI. This is a meaningful threshold for robotics and autonomous vehicle R&D teams who need production-grade tooling. Effective robot learning requires simulation that responds to behavior, not just renders a static scene. Oasis 3’s action-conditioned world generation lets a robot policy act inside an environment, receive feedback, and experience how the world changes in real time.
Powered by DOS 2.0, Decart’s purpose-built Optimization Stack, Oasis 3 delivers low-latency generation and interaction built for real development workflows. Simulation becomes an interactive learning environment, not a fixed asset.
Building Oasis 3 required tight control over a complex training workflow
Training a world model means juggling an unusual number of moving parts from visual fidelity, physical plausibility, action conditioning, and multi-view consistency all improving (or regressing) independently. Previously, comparing runs meant scattered logs and tribal knowledge.
W&B Models was the backbone of Oasis 3’s development. Every training experiment, from early architecture explorations to final large-scale runs, was tracked in one place. Sweeps accelerated hyperparameter optimization, while Artifacts versioned datasets, checkpoints, and model outputs so any result could be traced to the exact data and configuration that produced it. W&B Models’ lineage tracking and access controls deliver that audit trail by default. This makes provenance a property of the workflow, not a scramble at release time. For Decart and the robotics teams building on Oasis 3, that’s real peace of mind.
Knowing exactly which data, code, and checkpoints produced the model we serve is table stakes for working with serious robotics partners. W&B’s lineage tracking and access controls provide that audit trail by default, so when partners ask, the answer is already there. That’s what allows Decart to move fast on releases without compromising the compliance bar their customers require.
Real-time interactive world generation demands infrastructure built for it
A powerful model alone isn’t enough. Real-time interactive world generation requires a full optimization stack designed for speed, scale, and responsiveness with infrastructure that can match it. Decart’s DOS 2.0 runs on CoreWeave’s purpose-built AI cloud “and is built on top of NVIDIA HGX B200 systems, with CUDA libraries including cuBLAS and cutlass for optimizations. The model was trained using the NVIDIA Physical AI Open Dataset on Hugging Face, supporting the performance needed for real-time simulation workflows.” As Orian Leitersdorf, Chief Scientist at Decart, puts it:
“At Decart, we transform video from a static medium into a living, responsive world experience. Our partnership with CoreWeave has enabled us to run NVIDIA HGX B200 at production scale and push the boundaries of training and inference, while delivering seamless interactive video experiences via the CoreWeave stack. With the NVIDIA HGX B300 and the upcoming NVIDIA Vera Rubin platform, we are effectively enabling a new level of massively scalable real-time generative and agentic AI.”
Explore Oasis 3
Oasis 3 is available today via API. Whether you’re building robotic training pipelines, autonomous vehicle simulations, or physical AI applications at scale, discover the first API-accessible world model for physical AI and start building today.