Decart has launched Oasis 3, a real-time generative world model designed to synthesise photorealistic driving environments for autonomous vehicle testing, according to TechCrunch reporting published on 10 June 2026. The model is available immediately via API, positioning it as an infrastructure-layer tool for AV developers rather than a consumer-facing product.

What Oasis 3 Does

Oasis 3 generates extended, photorealistic simulations of driving scenarios in real time. The stated capability — hours of continuous, coherent driving environment synthesis — addresses one of the more stubborn bottlenecks in AV development: the cost and logistical difficulty of accumulating sufficiently diverse real-world driving data at scale. Synthetic data pipelines are not new to the AV stack, but the combination of real-time throughput and photorealism in a single generative model, delivered over an API, collapses what has traditionally required a bespoke simulation platform into something closer to a commodity service.

The TechCrunch piece notes caveats alongside the capability claims — a detail worth holding onto, and one this article returns to below.

Oasis 3 in the Context of Decart's World-Model Trajectory

Oasis 3 is the third major iteration of a product line that began with a demonstrably different application. The original Oasis, as described on Decart's site, is a real-time interactive generative video model trained on gameplay clips and keyboard inputs — in essence, an AI system that renders a playable game environment frame-by-frame from a neural model rather than a conventional game engine. Oasis 2.0 extended that concept with a video-to-video (V2V) architecture, transforming live game worlds as players interact with them, again in real time.

The pivot from interactive game simulation to autonomous driving simulation is meaningful. Both domains require real-time world-state consistency — objects must persist and behave physically plausibly across frames — but AV simulation adds strict demands around geometric accuracy, sensor-realistic rendering (cameras, LiDAR-adjacent appearance models), and edge-case density. Decart's decision to target the AV market with Oasis 3 implies the underlying world model has reached a fidelity and temporal coherence threshold that the gaming experiments, however attention-grabbing, were always functioning as a proving ground for.

Decart's broader model portfolio, which includes Mirage for environment generation and Lucy for human-figure synthesis, has also been integrated with XR headsets, per the company's developer documentation published in October 2025. The connective tissue across these products is a real-time inference architecture capable of maintaining world-state coherence at interactive frame rates — a non-trivial systems problem that sits well below the model architecture layer and tends to be underweighted in coverage of generative AI.

Why AV Simulation, Why Now

The autonomous vehicle industry has spent the better part of a decade wrestling with the simulation-to-reality gap. Deterministic simulation engines — think CARLA, LGSVL, or the in-house platforms run by Waymo, Cruise, and their peers — offer controllability and repeatability, but their rendering pipelines produce environments that trained neural perception stacks can distinguish from real sensor data, introducing distributional shift at inference time. Domain randomisation and neural rendering have both been applied as partial remedies, but neither fully closes the gap without significant engineering overhead.

Generative world models trained on real driving footage occupy a different position in that design space. Because they learn the statistical structure of real-world appearance directly, their outputs can, in principle, be closer to the true data distribution that a deployed perception model will encounter. The "photorealistic" claim in Decart's positioning speaks directly to this problem.

Worth flagging: the TechCrunch report specifically notes caveats around Oasis 3's capabilities. Without the full technical disclosure — Decart has not, as of this writing, published a detailed model card or peer-reviewed evaluation — it is not yet possible to independently characterise where the system's temporal consistency degrades, how it handles rare-event scenarios (the long-tail problem that is arguably the hardest part of AV validation), or what its failure modes look like when driven out of distribution. AV engineers evaluating this as a data-generation tool will want those specifics before integrating it into a validation pipeline with safety implications.

The API Delivery Model

The decision to release Oasis 3 via API rather than as an on-premises or self-hosted package has direct consequences for who can use it and how. For smaller AV teams and research groups without the infrastructure to run large generative models in-house, API access lowers the barrier considerably. For enterprise AV programs with data-sovereignty requirements — particularly those operating under automotive-sector regulations or defence contracts — sending scenario data to a third-party endpoint introduces compliance considerations that will need careful evaluation.

The API model also means Decart retains control over the model weights and, implicitly, over the scenarios customers can generate. That is a competitive and regulatory double-edge: it simplifies Decart's ability to iterate and patch the model, but it introduces dependency risk for customers building long-horizon validation programmes around a vendor-controlled service.

We have seen this dynamic play out before in cloud-era enterprise software: the shift from on-premises Oracle and SAP deployments to SaaS platforms in the 2010s produced exactly this tension between accessibility and control. Many enterprises eventually negotiated hybrid arrangements or multi-vendor strategies to manage it. AV programmes integrating Oasis 3 are likely to face similar architectural decisions sooner rather than later.

What Changes If This Works

If Oasis 3 delivers on its headline capabilities — and the caveats in early coverage suggest maturity is still accruing — the implications for AV development economics are worth thinking through carefully. Simulation hours are currently expensive to generate at the fidelity required for perception-stack training and validation. A generative API that can produce photorealistic, behaviourally coherent driving scenarios at scale would compress that cost curve in ways that could be significant for teams outside the tier-one AV players who can fund their own simulation infrastructure.

It would also shift some leverage in the AV toolchain toward the world-model layer — a layer that, until recently, has been largely owned in-house by the major programmes. Third-party generative simulation infrastructure could, over time, disaggregate what has been a vertically integrated stack.

Oasis 3 is early in that story. The API is live, the capability claims are substantive, and the caveats are real. The next meaningful data point will be independent technical evaluation — either from researchers, from early API customers publishing their findings, or from Decart's own forthcoming technical documentation. That is the evidence that will determine whether this iteration moves the needle on one of autonomous driving's longest-standing engineering constraints.