WeRide Launches Physical AI Foundation Model WITT, Processing 10,000 Minutes of Video Per GPU Per Day

WeRide officially launched WITT (World Intelligence Toward Truth), a physical AI cognitive foundation model, on July 17, targeting a core bottleneck in autonomous driving R&D: how to extract high-value long-tail scenarios from massive video datasets.

Built on a vision-language model (VLM) architecture, WITT introduces the concept of "minimum physical fact units," decomposing continuous real-world driving scenes into identifiable and verifiable factual elements. The model's name pays homage to philosopher Ludwig Wittgenstein's idea that "the world is the totality of facts."

WITT comprises four core capabilities: fact extraction, fact reasoning, fact verification, and fact orchestration. Fact extraction identifies driving behaviors, multi-agent interactions, and physical conditions from video. Fact reasoning analyzes causal relationships and how events evolve. Fact verification evaluates outputs across six dimensions through a mechanism WITT calls "6+1 Verification," while fact orchestration routes data to simulation, training, or human review based on scarcity and learning value.

According to WeRide, WITT reduces token costs by 98% compared to hundred-billion-parameter general models, processes 10,000 minutes of driving video per GPU per day, and achieves up to 200x efficiency improvements. Its per-segment factual error rate in autonomous driving scenarios is roughly one-third that of general-purpose large models.

WITT operates on the cloud alongside WeRide's previously released world model GENESIS, forming a "physical AI flywheel." WITT extracts and verifies facts from real-world L4 fleet data, while GENESIS generates high-fidelity simulation scenarios and long-tail variants for training both L4 and L2++ vehicle models.

WeRide's L4 autonomous driving fleet now exceeds 3,000 vehicles, with Robotaxi services operating fully driverlessly in Guangzhou, Beijing, Abu Dhabi, and Dubai. Its L2++ end-to-end solution WRD 3.0 has secured nearly 30 vehicle model programs, with production deployments on vehicles from Chery, GAC Aion, and others across China, Germany, France, and Japan.

WITT's arrival signals a shift from data quantity to data quality in autonomous driving development. As fleet data continues to accumulate, the ability to efficiently extract high-value long-tail scenarios will increasingly determine the pace of model iteration.