Key takeaways
- Qwen-Drive-1.0-4B was open-sourced on 6 September 2026. Built on Qwen3.5-4B, it is 4 billion parameters.
- Its significance is not size but consolidation: 3D perception, visual question answering and motion planning in a single model — three separate systems, three codebases and three calibration datasets, previously.
- Weights are downloadable from GitHub, Hugging Face and ModelScope, free.
- The practical consequence: a newly founded carmaker, or a small team doing autonomous haulage in a mine, no longer has to accumulate data and train a model from zero.
The question at the dealership
Last weekend I went to a car dealership with a friend. The salesperson tapped through the centre display and quoted a price for the intelligent driving package: RMB 38,000.
My friend turned and asked what exactly that bought.
I did not have an answer. What happened on 6 September may reprice that sticker.
Why a four-billion-parameter model makes carmakers nervous
The model is small. That is the point.
Until now, autonomous driving stacks were built as separate modules: one that perceives what is on the road, one that answers what a thing is, one that decides how to drive next. Three teams, three codebases, three sets of calibration data — and a great deal of engineering spent making them agree with each other.
Qwen-Drive folds those into one model, small enough to fit on vehicle-side compute.
Then there is the word "open." The weights are on public repositories. Anyone can download them.
That changes who gets to compete. A carmaker founded last year, or a small team building autonomous vehicles for a mine site, does not have to start by assembling a data operation. The barrier drops by a large step.
Once a technology is free, the premium charged for it starts counting down.
So will the option price fall?
Not tomorrow. But the direction has changed.
- Mid and low-end cars get usable driver assistance. Previously a feature discussed only above RMB 200,000; now it fits into a car costing just over RMB 100,000.
- The option package shifts from selling software to selling standard equipment. When a competitor makes it standard, charging RMB 38,000 separately stops working.
- Used-car residual values for driver assistance get re-rated. The premium you pay today may return nothing when you sell in three years.
There is a counterintuitive conclusion here: the later you buy, the better the deal. That is not advice against buying a car. It is advice against paying a large standalone premium for the word "intelligent."
What to watch instead of the price
One thing has not changed. On every driver-assistance system in mass production today, the responsible party is still the person in the driver's seat.
Open-sourcing spreads capability. It does not spread liability. If the system mistakes a plastic bag for a rock, or misjudges a cyclist in a raincoat, the answer to "who bears it" has not moved.
And this class of model has a built-in weakness: it recognises what it has seen, and hesitates on what it has not. Extreme weather, oddly shaped obstacles, temporary roadworks — the long tail is where safety actually lives.
What to do now
If you do not need a car urgently, wait a year. The price curve for driver assistance is falling steeply, and buying a year early can cost you two years of depreciation on that feature.
At the dealership, ask about delivery, not roadmap. One question filters out most of the storytelling: does this function work today, or does it arrive in a later OTA update?
And keep the word "assistance" in mind. As long as the steering wheel is in your hands, the responsibility is yours. Looking away at any moment is betting your family's safety on a probability.
What RMB 38,000 bought was never the technology. It was the reassurance that other people had it. Once the technology is free, that reassurance depreciates fast.
Honest limitations
"First vision-language foundation model for autonomous driving" is the developer's characterisation. Open-weight availability does not make a model road-legal; vehicle integration, validation and regulatory approval remain the dominant costs and are not addressed by a model release. Pricing effects described here are reasoning, not observed data.
Sources: Qwen open-source release (6 September 2026); model repositories. Information only.