DeepRoute.ai presents 40B VLA model for scalable autonomy

DeepRoute.ai has presented its 40-billion-parameter vision-language-action (VLA) foundation model at Nvidia GTC 2026, describing an architecture that compresses the autonomous driving data processing cycle from more than five days per iteration to approximately 12 hours. The company has deployed its autonomous driving systems across more than 250,000 production vehicles and is targeting one million vehicles by the end of 2026.

The foundation model performs three simultaneous functions: acting as the driver by executing real-time driving actions, as an analyst by identifying critical events and providing causal reasoning, and as a critic by evaluating trajectories for safety, comfort and naturalness. By embedding all three roles within a single model, DeepRoute.ai has automated large portions of the data pipeline, enabling autonomous identification of high-value training scenarios such as near-misses, root-cause analysis and annotation generation without manual intervention.

The architecture creates a self-reinforcing development cycle in which improvements in driving performance directly enhance the system’s ability to curate its own training data, reducing dependence on manual labelling. DeepRoute.ai captured approximately 40% market share among third-party suppliers in the high-level autonomous driving segment in October 2025.

In a statement, Tongyi Cao, Chief Technology Officer of DeepRoute.ai, said: “The bottleneck is no longer about acquiring data; it is about how efficiently a system can filter out the noise and convert massive amounts of raw data into high-value training samples. By leveraging our Foundation Model, we have entirely reconstructed this workflow.”

Source: PR Newswire