Decart · 2026-06-10 · major
Decart Ships Oasis 3 — First API-Accessible Interactive World Model for Physical AI Streams Three Synchronized 768×512 Camera Views at 22 FPS With <200ms End-to-End Latency on NVIDIA HGX B200, Priced at $0.02 per Second of Simulation
Decart's flagship world model generates infinite photorealistic driving scenes from a text prompt, takes throttle/steering actions as input, and returns three synchronized camera views — available now via gRPC API for autonomous-vehicle and robotics training.
Decart's Oasis 3 lets robotics and AV teams stream three synchronized photorealistic camera views in real time from a text prompt, action-conditioned, via API.
Key specs
| Price | $0.02 per second |
|---|---|
| Fps | 22 |
| Latency | <200ms end-to-end |
| Resolution | 512×768×3 |
| Camera views | 3 (left_forward, front, right_forward) |
| Recent raise | $300M (May 2026) |
What is it?
Oasis 3 is an interactive world model: feed it a scene description plus throttle and steering actions and it streams three synchronized 768×512 camera views (left-forward, front, right-forward) at 22 FPS for as long as you want to drive. It is action-conditioned, multi-view consistent, and exposed as a gRPC over HTTP/2 API with a 120-second per-RPC timeout. Decart positions it as the first API-accessible world model purpose-built for physical AI rather than a research demo.
How does it work?
Each call accepts four [throttle, steering] pairs in [-1, 1] and a scene prompt; the server returns four RGB frames per camera view as VP9-encoded streams that the SDK decodes to uint8. Generation is autoregressive 'Live Stream Diffusion' on the Decart Optimization Stack (DOS), trained on the NVIDIA Physical AI Open Dataset and served on NVIDIA HGX B200 systems via CoreWeave. End-to-end latency stays under 200ms so the world reacts inside a real training loop.
Why does it matter?
Until now world models like Genie 3 and World Labs Marble were either offline batch demos or single-view toys; Decart claims order-of-magnitude lower per-second cost and the first API path for closed-loop reinforcement learning, policy evaluation before real-world deployment, and high-fidelity teleoperation. TechCrunch's hands-on flagged real failure modes — scenes degrade over long sessions and physics can let cars phase through obstacles — so the bar is set for rivals, not cleared.
Who is it for?
Autonomous-vehicle teams, robotics labs, simulation engineers building closed-loop RL pipelines.
Try it
DECART_API_KEY=… ; model: oasis-3-preview ; docs: docs.platform.decart.ai/models/realtime/oasis-3