The autonomous driving industry has seen no shortage of the term "Physical AI" this year. From the GTC conference in early 2026 to the Beijing Auto Show, nearly every company has been discussing large models and foundational models, yet most of their products still operate on small models.
Yuanrong Qixing CEO Zhou Guang offered a sharper assessment at the Future Automotive Pioneers Conference on May 29: "The difference between one human intervention every few dozen kilometers and one every thousand kilometers represents two entirely different species."
This statement comes against the backdrop of Yuanrong achieving a year-on-year growth rate of 2.1 times in the third-party urban NOA segment in 2025, with its market share briefly reaching 38% in October alone. The target for 2026 is to deliver over one million units and surpass a thousand kilometers in urban scenario MPCI (Miles Per Critical Intervention).
Behind this ambition is Yuanrong's comprehensive shift from small models to large models.
Zhou Guang used an internal analogy to explain the bottleneck of small models: the "seesaw effect."
When Version A is specially tuned for Shanghai and Wuhan, performance in Shenzhen, Guangzhou, and Beijing deteriorates. The next version fixes Beijing, but issues arise in mountainous road scenarios. This constant back-and-forth between versions leads users to perceive the system as inconsistent, making it difficult to build trust.
Over the past five years, the entire industry has evolved from multi-module stitching to end-to-end systems and then to the consolidation of small models, with technological iterations largely confined to the same framework. Zhou Guang believes this approach has reached its peak, where "investment increases while improvements slow down."
Yuanrong's proposed solution is a 400-billion-parameter VLA (Vision-Language-Action) foundational model that integrates three capabilities into a single model: Driver for driving decisions, Analyst for scene understanding and data annotation, and Critic for evaluating the quality of driving behavior.
Zhou Guang emphasized that a large model is not merely a tenfold scaling of a small model's parameters. He provided an example: if a dog is painted with zebra stripes, a small model might identify it as a zebra based on the stripes, while a large model, considering the overall form, would correctly identify it as a dog.
In essence, small models rely on local features for conditioned responses, whereas large models make judgments based on holistic cognition. In long-tail scenarios for autonomous driving, such as construction detours, irregular obstacles, or unmarked intersections, the gap between the two becomes significant.
Yuanrong claims that after introducing the foundational model, many labor-intensive steps in the traditional data closed-loop have been handed over to the model, improving data efficiency tenfold.
For 2026, Zhou Guang has set the targets at one million deliveries and a thousand kilometers for urban MPCI. Achieving an average of one human intervention per thousand kilometers on urban roads would be over ten times the current industry mainstream level. He believes that at this safety threshold, daily user adoption could exceed 50%, making subscription-based payment models viable. The reference point is Tesla's V14, which, with a monthly fee of $100 in the U.S. market, already sees 50% user adoption.
His stance on Tesla's FSD entering China is clear: "This is a positive signal." Just as Tesla once spurred China's electrification wave, the entry of FSD will redefine industry standards for safe autonomous driving. For companies already invested in large models, this represents an opportunity to reallocate market share as standards are raised; for those still operating within the small-model framework, "it could be quite challenging."
However, the thousand-kilometer MPCI target remains a goal, not an achievement. Yuanrong's 40B foundational model was first unveiled at GTC in March this year, with a technical breakdown presented by former DeepSeek R&D lead and current Chief Scientist Ruan Chong at the Beijing Auto Show. There is still a distance to go before mass production and vehicle integration. Zhou Guang himself acknowledged, "Today's autonomous driving is still quite homogenous in competition," noting that end-to-end capability is no longer exclusive to one or two companies.
In the interview, he offered a more specific vision for the industry's endgame: the final race will not be about "how many autonomous driving companies remain," but rather three types of players—large model companies entering physical AI, autonomous driving companies developing their own large models, and automakers building their own large models. Each direction will likely sustain one or two players, totaling three to five major contenders. Zhou Guang is betting that Yuanrong can secure a position in the second category.
Comments