DeepRoute.ai Abandons Small Models, Bets on Physical AI to Reach 1,000-Kilometer MPCI

Introduction: A Strategic Divergence in the Intelligent Driving Industry

While most autonomous driving companies are still competing over how many cities they've launched in or how small and fast their models are, DeepRoute.ai has chosen a radically different path — abandoning the small model strategy, opting out of the city-launch war, and instead betting on a physical AI foundation model with the goal of reaching 1,000-kilometer MPCI (Miles Per Critical Intervention).

This strategic pivot is not only a major adjustment to DeepRoute.ai's own technical roadmap but also reflects a deeper transformation underway across the intelligent driving industry: a shift from "scale expansion" to "capability leaps," and from "stacking city counts" to "building systems that truly work."

The Core Question: Why Abandon Small Models?

Over the past two years, end-to-end small models were widely regarded as the "optimal solution" for mass-produced intelligent driving. Companies raced to release lightweight models with fewer parameters, faster inference, and lower deployment costs, attempting to deliver decent driving performance within the limited computing power available on vehicles. However, DeepRoute.ai CEO Zhou Guang recently stated explicitly that the company has decided to move away from the small model approach.

The reasoning is straightforward. While small models have advantages in deployment efficiency, their capability ceiling is equally apparent. When confronted with complex long-tail scenarios, small models often lack sufficient generalization ability and reasoning depth, resulting in systems that require frequent human intervention at critical moments. In other words, small models can make intelligent driving "run," but they struggle to make it "run far."

DeepRoute.ai's assessment is that the next key metric in intelligent driving competition is not "how many cities are covered" but "how far the system can travel between critical interventions" — that is, MPCI. Only when this number breaks through the 1,000-kilometer mark or higher can intelligent driving truly transition from "assisted driving" to "autonomous driving" and establish genuine commercial value and user trust.

Physical AI Foundation Model: The Weapon for a Different Battle

After abandoning small models, DeepRoute.ai has shifted its technical focus to a "physical AI foundation model." Physical AI refers to artificial intelligence systems capable of deeply understanding and modeling the laws of the real physical world. Unlike traditional perception-planning-control separated architectures, the physical AI foundation model attempts to build a unified large model with deep comprehension of the physical world.

Such a model must not only "see" vehicles and pedestrians on the road but also "understand" their movement intentions, predict their behavioral evolution under different physical constraints, and even reason about possible reactions of traffic participants in extreme scenarios never previously encountered. This is precisely the capability that small models lack.

From a technical implementation perspective, a physical AI foundation model requires larger parameter counts, richer training data, and stronger reasoning capabilities. This means higher R&D investment and computing power demands, but it also means a higher capability ceiling. DeepRoute.ai clearly believes that as intelligent driving technology matures, the capability ceiling matters more than deployment cost.

Notably, DeepRoute.ai also announced it will not participate in the "city-launch race." Over the past year, many intelligent driving companies have used "number of cities launched" as a core marketing metric, frequently announcing coverage of hundreds of cities. However, industry insiders widely acknowledge that many of these so-called city launches are more marketing than substance, with actual usability and safety varying greatly. DeepRoute.ai has chosen to step out of this competitive dimension entirely, concentrating resources on improving the system's core driving capabilities.

Analysis: What Does 1,000-Kilometer MPCI Mean?

Currently, publicly available MPCI data for intelligent driving systems in the industry generally ranges from tens to hundreds of kilometers. This means drivers need to perform a critical intervention every few tens to few hundred kilometers on average to avoid potential safety risks.

If DeepRoute.ai can push MPCI to the 1,000-kilometer level, it would represent a qualitative leap. A 1,000-kilometer MPCI means that a driver traveling from Shenzhen to Beijing — a journey of approximately 2,000 kilometers — would theoretically need only one to two critical interventions for the entire trip. Such performance would approach or even surpass the level of an average human driver.

Of course, achieving this goal faces enormous challenges. The physical AI foundation model demands far more computing power than small models, and how to efficiently deploy it on vehicles with limited computational resources is a key problem. Additionally, the cost of collecting and annotating the high-quality data needed to train a physical AI model should not be underestimated.

From a commercial perspective, DeepRoute.ai's strategy also carries certain risks. In the current industry trend of cost reduction and efficiency improvement, choosing a heavy-investment foundation model approach requires sufficient financial backing and patience. If the company cannot produce convincing results in the short term, it may face dual pressure from investors and the market.

However, from another angle, precisely because most companies are pursuing the small model and city-launch approach, DeepRoute.ai's differentiated strategy could form a unique competitive moat. Once the physical AI foundation model's capabilities are validated, its technological barrier will be far higher than that of competitors relying on lightweight models.

Outlook: The Second Half of the Intelligent Driving Race

DeepRoute.ai's strategic pivot may signal that the intelligent driving industry is entering its second half. The first half of the competition centered on "whether you have it" — whether you have an end-to-end model, whether you've launched in enough cities, whether you've achieved mass production. The second half will center on "whether it's good enough" — whether the system can actually operate safely and reliably over extended periods in real-world scenarios.

The concept of physical AI has attracted increasing global attention in recent years. NVIDIA CEO Jensen Huang has repeatedly emphasized in public speeches that physical AI will be at the core of the next technological wave. DeepRoute.ai's decision to adopt physical AI as its core technology pathway in autonomous driving demonstrates its forward-looking judgment on industry trends.

In the future, as vehicle-side chip computing power continues to increase and cloud-based training infrastructure continues to improve, the conditions for deploying physical AI foundation models will gradually mature. Whether DeepRoute.ai can use this approach to be the first to break through the 1,000-kilometer MPCI milestone will be the key measure of its strategic success.

The endgame of the intelligent driving industry may not be about who has launched in the most cities, but about whose system truly earns users' trust to hand over the steering wheel. DeepRoute.ai has chosen a harder but potentially more correct path. Now it all comes down to execution.