Changan ships a self-developed end-to-end driving LLM after 1.45 billion km of validation
Changan became the latest major Chinese automaker to put a self-developed end-to-end driving model on the road, releasing what it calls an in-house large model for assisted driving that has been validated on roughly 1.45 billion kilometers of driving data, according to Chinese financial outlets. The release puts one of China's "big four" state-owned automakers into the same architectural camp as XPeng, Huawei, Li Auto and the other Chinese OEMs that have decided the road to L3 and beyond runs through one model that maps sensors straight to trajectories, instead of a hand-stitched stack of perception, prediction, planning and control modules.
Changan's model is the culmination of a multi-year build-out that began with rule-based driver-assist systems and migrated through increasingly large neural networks; the 1.45 billion kilometer validation figure is a cumulative mileage across the fleet used to train and test the system, not a per-vehicle number, and is meant to put the model within shouting distance of the data scale that Tesla, Huawei and XPeng have been accumulating. Like the rest of the field, Changan's system is generative: a vision-language-action style architecture that takes camera, lidar and map inputs and emits driving decisions token-by-token, the same broad approach that has powered the recent wave of "VLA" and end-to-end launches across the Chinese auto sector.
The commercial logic is the same as everywhere else in the industry: whoever owns the model owns the data flywheel, and whoever owns the data flywheel builds the brand. End-to-end models get better as their fleet drives more kilometers, which makes the model itself a procurement story as much as a technical one — buyers are increasingly picking their next car on the basis of which company's driving assistant will improve faster over the life of the vehicle. Changan's bet is that a state-owned OEM with a multi-million-vehicle annual sales base can out-accumulate smaller pure-play rivals on data, even if it started later on the model architecture itself.
The risk is the usual one. End-to-end systems are notoriously hard to evaluate pre-deployment, and a model that performs well on 1.45 billion kilometers of aggregate fleet data can still fail in distribution shifts the training set never saw — construction zones with hand signals, police directing traffic, the kind of edge conditions that have already tripped Tesla FSD, Cruise and Huawei ADS in widely-circulated videos. Changan's release says nothing about formal regulatory approval for L3 or higher, and Chinese regulators have not yet blessed any production car for hands-off, eyes-off operation outside narrow geofenced trials. The model is shipping as a smarter L2+ system first, with the L3 conversation to follow.
What to watch: whether Changan publishes a benchmark comparing its model to XPeng's VLA 630, Huawei's ADS 4 or Li Auto's MindVLA on a standardized eval, and how fast it pushes the model to its Deepal, Avatr and Changan-branded passenger lines — the 1.45 billion kilometer figure only matters if it translates to shipped, dealer-floor product within a quarter, not a roadmap slide.
Do you think a state-owned automaker can out-iterate the pure-play Chinese AI-driving labs on data alone? Tell us in the comments.
