Decart's new world model can simulate hours of photorealistic driving - with some caveats

Overview

Decart, a two-year-old AI startup, released Oasis 3, a real-time world model that builds photorealistic driving scenes for testing self-driving cars. The launch follows a $300 million funding round that values the company at close to $4 billion, with strategic money from Toyota, Adobe, eBay, and Nvidia. The model is open to developers through an API at $0.02 per second, though early hands-on testing by TechCrunch found the simulation still breaks down on physics and scene consistency.

Key Takeaways

Oasis 3 generates interactive, photorealistic driving environments in real time, aimed at letting autonomous vehicle teams rehearse rare and dangerous situations without driving real cars.
Decart raised $300 million at a valuation near $4 billion, backed by Toyota, Adobe, eBay, and existing investor Nvidia.
The model runs through an API at $0.02 per second, with separate enterprise pricing for larger workloads.
Decart positions Oasis 3 as infrastructure other companies build on top of, rather than a finished simulation platform it sells end to end.
TechCrunch testing showed clear limits: cars pass through each other, scenes lose consistency over long sessions, and steering controls sometimes stop responding.
Decart says it has spent well under $100 million across its life, crediting in-house software that runs its models at lower cost than rivals.

Stats & Key Facts

#$300 million raised in the latest funding round
#Valuation of nearly $4 billion after the raise
#Roughly 8,000 tokens used per generated frame
#Tens of frames per second, adding up to hundreds of thousands of tokens per second
#$0.02 per second API price for generated simulation
#More than 100,000 developers in Decart's community around its Lucy video model
#Lifetime company spending kept well under $100 million
#Two years old as a company

What Oasis 3 Does for Self-Driving Car Testing

The product targets one of the hardest problems in autonomous driving: practicing for events that almost never happen on real roads.

Oasis 3 is an interactive world model, meaning software that builds a living, navigable environment rather than a fixed video clip. It produces photorealistic driving scenes that respond as a user or system steers through them, so an autonomous vehicle team trains and tests its perception software against situations it controls.

The appeal is edge cases. A child running into the street, a sudden swerve, an unusual road layout, these are dangerous and rare to capture by driving real cars. A world model lets engineers spin up those moments on demand and repeat them as often as needed, which is where Decart aims the tool.

The $300 Million Raise and Near $4 Billion Valuation

The release arrives alongside a large funding round that places Decart among the higher-valued startups in the world model field.

›Decart raised $300 million in its latest round.
›The round values the company at nearly $4 billion.
›Strategic investors include Toyota, Adobe, and eBay.
›Nvidia, an existing backer, also took part.
›Toyota's involvement ties the money directly to the automotive testing use case.

How the Model Generates Frames One at a Time

Under the hood, Oasis 3 uses an auto-regressive design, building each frame in sequence from the ones before it.

The system generates video one frame at a time and supports a multi-camera setup with one front-facing and two side-facing views, matching how a real vehicle senses its surroundings. It produces tens of frames per second and uses roughly 8,000 tokens for each frame, which works out to hundreds of thousands of tokens every second.

Oasis 3 is built on Lucy, Decart's real-time video foundation model, the same technology behind a developer community of more than 100,000 users. A key difference from rival research previews is duration: Oasis 3 runs a session for hours rather than seconds, which matters for sustained driving tests.

Where the Simulation Still Falls Short

TechCrunch ran the model directly and reported several caveats that buyers should weigh.

›Physics break down, with vehicles driving straight through other vehicles.
›Scene continuity degrades over time, so starting locations vanish and the world drifts into nonsensical, dream-like results.
›Controls over vehicle direction stopped responding at points during testing.
›Opening scenes look highly photorealistic but lose quality across longer sessions.

The Decart Optimization Stack and Low Spend

Decart credits its cost advantage to in-house software rather than raw compute spending.

The company points to the Decart Optimization Stack, software it built to run its models efficiently across hardware. By its own account, the company describes the stack as optimizing performance across Nvidia, Amazon, and Google hardware, letting the models run at lower cost than competitors.

That efficiency claim is tied to a striking number: Decart says it has spent well under $100 million over its entire life. For a company now valued near $4 billion and shipping a real-time world model, that lean track record is part of its pitch to investors and customers.

How Decart Sits Against Genie 3, Marble, and Others

Oasis 3 enters a crowded race to build usable world models.

›Google's Genie 3 is a direct rival in interactive world generation.
›World Labs offers Marble in the same space.
›Luma and Runway bring competing video-generation work.
›Decart frames Oasis 3 as infrastructure developers program on top of, not a finished end-to-end platform.
›CEO Dean Leitersdorf describes it as the first usable world model people will build on.

Frequently Asked Questions

What is a world model in plain terms?

A world model is software that builds an interactive, navigable environment rather than a fixed video. Oasis 3 generates photorealistic driving scenes that respond in real time as a user or system steers through them.

Who is Oasis 3 built for?

It targets autonomous vehicle companies that need to test and train their software against rare and dangerous driving situations. Generating those scenes on demand avoids the cost and risk of capturing them with real cars on real roads.

How much did Decart raise and at what valuation?

Decart raised $300 million in its latest round, valuing the company at nearly $4 billion. Backers include Toyota, Adobe, eBay, and existing investor Nvidia.

What are the main weaknesses found in testing?

TechCrunch reported that physics break down, with cars passing through each other, and that scenes lose consistency over longer sessions. Direction controls also stopped responding at times during testing.

How do developers access Oasis 3?

It is available through an API priced at $0.02 per second, with separate custom pricing for larger enterprise use. Decart presents it as infrastructure other companies build their own tools on top of.

Oasis 3 shows how far real-time world models have come for self-driving research, paired with a fresh $300 million in funding and a near $4 billion valuation. The early caveats around physics and consistency are a reminder that the technology is promising but not yet dependable for production testing.