DeepRoute.ai CEO Zhou Guang: ADAS Must Shift to Large Model Paradigm to Boost Data Efficiency and Safety

At the 4th Future Mobility Pioneers Conference, Zhou Guang, Founder and CEO of DeepRoute.ai, delivered a speech highlighting that over the past five years, advanced driver-assistance systems (ADAS) have primarily relied on small models, whose performance improvements have significantly slowed due to approaching their technical limits. He argued that achieving a substantial leap in ADAS capabilities requires a paradigm shift toward cognition-based technologies represented by large models. Zhou explained that while small models excel at reflexive responses and local feature recognition, large models demonstrate clear advantages in holistic perception and higher-level reasoning. He proposed integrating three core capabilities—Driver (driving execution), Analyst (data analysis), and Critic (system evaluation)—into a unified foundation model. This integration, he said, would create a data flywheel effect across development, iteration, and operations, boosting data efficiency tenfold and reducing reliance on manual processes. DeepRoute.ai has already deployed a 40-billion-parameter foundation model to support dual scaling in both models and data. This model not only powers on-vehicle decision-making but also enhances data mining and system evaluation. According to disclosed figures, DeepRoute.ai captured a 24% market share in third-party Navigation on Autopilot (NOA) solutions in 2025, ranking second in the industry. Its year-over-year market share growth of 2.1x was the highest in the sector, with one out of every three newly sold vehicles equipped with NOA adopting DeepRoute.ai’s solution. The company plans to deliver over one million vehicles in 2026 and aims to increase the Mean Miles Between Disengagements (MPCI) on urban roads to over 1,000 kilometers through its foundation model, targeting a user adoption rate exceeding 50% for frequent usage. Additionally, since late 2024, DeepRoute.ai has focused on Physical AI, officially positioning itself as a Physical AI company in March 2025 and subsequently showcasing related technological advancements at industry events.