Nvidia Just Released a Free Self-Driving Brain Small Enough to Run Inside a Car

Every carmaker that wants a car to drive itself runs into the same wall. The smart software that can actually reason about a messy street is huge, and the computer bolted into a car is not. You could have a genius in a data center or a quick thinker in the dashboard, not both.

Nvidia says it just closed a big piece of that gap, and it isn’t charging for the software. On Tuesday the company posted Alpamayo 2 Edge on Hugging Face, a compact driving model built to run inside the car on Nvidia’s DRIVE AGX Thor computer. It looks at the road through the car’s cameras, plans where to go next, and, when a situation gets tricky, writes out in plain words why it’s doing what it’s doing.

The weights are free to download under an open license, and the code is under Apache 2.0, the same permissive license a lot of everyday software uses. Any automaker, startup or university lab can grab it, test it and build on it.

Here is Nvidia’s own demo of it working through real traffic.

A driver that only talks when it has to

The clever part is how little it says. Earlier reasoning models for cars would narrate everything, and all that narration burns computing power the car doesn’t have to spare. Alpamayo 2 Edge reasons “only when needed,” in Nvidia’s words. On a boring stretch of highway it spits out a short internal tag that means, roughly, “nothing unusual here,” and saves its effort. When a crossing guard steps out or a bus swings wide, it starts explaining itself.

That explanation is more than a party trick. If a car brakes hard for no obvious reason, engineers want to know what it saw and what it was thinking. A model that writes down “slowing because the pedestrian on the right is about to step off the curb” is a model you can actually check.

Under the hood, the model takes in video from cameras all around the car, a short history of how the car has been moving, and plain navigation instructions like a turn ahead. From that it predicts 64 positions covering the next 6.4 seconds of driving. According to the model card, the core is about 2.45 billion parameters, plus a small 0.2-billion-parameter piece that turns its plan into a smooth path. Nvidia rounds the whole thing to 2 billion.

For comparison, its big sibling, Alpamayo 2 Super, has 34 billion. Nvidia used that larger model as a teacher, having it write explanations for real driving clips and training the small one to copy its thinking.

Fast enough for the dashboard

Speed is the whole ballgame in a moving car. Nvidia’s technical write-up says that once the model is compressed and tuned for the Thor chip, it goes from camera image to driving plan in 160 to 190 milliseconds, depending on how much reasoning a moment calls for. That’s under a fifth of a second, fast enough for the roughly 10 decisions a second Nvidia designed it for.

Nvidia published the full recipe for getting there, including the exact Thor board, operating system and software versions it measured on, so other developers can check the number for themselves instead of taking the company’s word for it.

Nvidia also trained it the way you’d train a teenage driver: by letting it make mistakes. In a simulator called AlpaGym, the model drove, saw the consequences of its choices and adjusted. Nvidia says that round of practice improved its closed-loop driving score by 48.9% over the best version trained without it.

The safety number Nvidia is leading with

The headline stat is distance between at-fault incidents, meaning how far the car goes in simulation before it causes a problem. Nvidia compared its model with Qwen-Drive-1.0, a 4-billion-parameter driving model, twice the size.

That number deserves a little care. It comes from Nvidia’s own simulation, not from cars on public roads, and Nvidia’s blog chart puts the gap at about 60 times while its 𝕏 post says 40. Either way it’s a lopsided result, but it is a company measurement in a company simulator, and independent tests on real streets will matter more.

Outside testers moved fast

The outside companies with early access tested it on their own data, which makes their results worth a look.

PORT, which records its own London street drives, ran the model in rebuilt simulations of those streets within days. Its chief technology officer, Matas Simonavicius, said that in early tests the car “kept its lane in more than twice as many runs” as Nvidia’s previous version, recovered from drifting out of its lane about 40% more often, and planned roughly twice as fast as their hand-tuned older setup on the same chip.

Niulinx, another early tester, said it had the model running its first tests on a real vehicle within a few days. Its demo shows the car stopping for a traffic officer’s hand signal and yielding to pedestrians in a crosswalk. Fine-tuning on Niulinx’s own driving data cut the model’s path-prediction error by another 44.2%, co-founder Luca Crotti said.

Embedl, a company that specializes in squeezing models onto small chips, said it had the model running at 10 times a second on Thor in less than 48 hours.

In Japan, TIER IV, the company behind the open-source Autoware driving software, announced Wednesday that it has finished plugging Alpamayo 2 Edge into Autoware and is now adapting it to real vehicles, according to its press release. Marco Pavone, the Nvidia researcher who co-wrote the technical post, cheered the speed of that work and said the model is already “experiencing the streets of Tokyo.”

Why giving it away makes sense for Nvidia

Nvidia doesn’t need to sell this model to make money from it. It sells the chips. Every carmaker that builds on a free, open driving brain tuned for Thor is a carmaker more likely to buy Thor, and Nvidia has spent the past year building out a full kit around it: the big teacher model, the small car-ready model, the simulator and a grading tool called AlpaJudge that scores how sensible a model’s explanations are.

None of this shows up in a showroom tomorrow. These are building blocks, and the companies using them still have to prove their systems safe. But the trend is clear. The kind of driving intelligence that used to need a rack of servers is starting to fit in the car itself, and it can tell you why it hit the brakes.