Aion N60 scores 117.81 to win Guiyang qualifier

On August 15, on the Guiyang course, the GAC Aion N60 scored 117.81 points among 18 participating vehicles, out of a maximum of 120, with zero interventions throughout the qualifier. The result at the Guiyang stop of the second intelligent-driving competition organized by First Electric Network created a striking contrast: a vehicle with a top-trim price in the 120,000-yuan range set a record since the competition began 30 stops ago, despite having unremarkable hardware—a Qualcomm 8650 domain controller, about 200 TOPS of computing power, one 192-line lidar and two 4D millimeter-wave radars. Its specifications were not exceptional among the participating vehicles.

In the past, people have tended to equate larger models, higher TOPS figures and lidars with more lines directly with stronger intelligent driving. The team behind the N60 offers a different answer: change how the resources are used. The end-to-end approach puts perception, prediction and planning into the same network, generating vehicle trajectories directly from sensor inputs. WeRide’s model for L4 Robotaxis was distilled and compressed into a smaller model suited to roughly 200 TOPS. A smaller model does not mean its capabilities disappear in the same proportion, but that does not mean a small model will outperform a large model running on a higher-computing-power platform.

Computing power is only a budget. How much of it is actually converted into driving capability also depends on model compression, operator support, memory access, heterogeneous scheduling and software deployment. After ultimately choosing the Qualcomm 8650, GAC worked with WeRide and Qualcomm to mobilize the NPU, CPU and GPU, pursuing a strategy of “using heavier software and engineering to trade for lower hardware costs.” Data is not just about volume, either: WeRide’s Robotaxi fleet is driven by the same system, producing more consistent behavior that is easier to process afterward. Extreme cases accumulated through a decade of operations can also be brought back into simulation through the Genesis cloud-based world model.

The competition did not erase every problem. The N60 had also failed at several previous stops, while its smaller body made it easier to maneuver through the three-point turns in Guiyang; live traffic and luck also played a role. The end-to-end approach is more composed in complex urban interactions and narrow roads affected by construction, but it is not necessarily as strong in highway scenarios. WeRide is currently addressing issues such as matching stop lines with traffic lights and merging into the target lane when turning right through continuous iteration. Version 2.3.0 has optimized highway driving, and the next version is expected to be released next month.

Specifically, consumers in the 100,000-yuan range are not getting a comprehensively stripped-down car costing more than 200,000 yuan; they are getting a set of trade-offs that concentrates the budget on intelligent driving. Supply-chain price reductions have lowered the costs of cameras, lidar, millimeter-wave radar and chip computing power per unit, while platformization allows the same sensors, domain controllers, software adaptations, testing and OTA capabilities to be reused across models. The N60 has accumulated nearly 20,000 deliveries, with monthly sales of about 5,000 units, and has not yet become a runaway hit. But it shows that intelligent-driving capability is starting to detach partly from vehicle price and factory equipment, and that the software version—not just the TOPS figure on the specification sheet—may be what truly determines the experience.