EVwire brief: NVIDIA has released Alpamayo 2 Edge, an open self-driving AI model small enough to run in real time on the company's DRIVE AGX Thor in-car computer. The model looks at a vehicle's surround cameras, plans the path ahead and explains its driving decisions in plain language.
Alpamayo 2 Edge only reasons out loud when a situation calls for it. During routine driving, it outputs a simple "nominal" signal instead of a full explanation, which saves computing power for the moments that matter, like yielding to pedestrians or making room for an oncoming bus.
With NVIDIA's optimizations, the 2-billion-parameter model goes from camera images to a planned trajectory in 160 to 190 milliseconds on DRIVE AGX Thor, depending on how much reasoning it does. It is available now on Hugging Face for developers to build on and adapt to their own fleets.
NVIDIA shared a video of Alpamayo 2 Edge in action on X:
Thinking of getting an EV, selling yours, or need EV accessories and adapters?
Check out our fresh EVwire EV Deals page for the discounts we've gathered.
How Alpamayo 2 Edge works
Alpamayo 2 Edge is a vision-language-action (VLA) model, meaning a single AI system sees the road, reasons about it in words and outputs a driving path. Here's what goes in and what comes out:
Inputs: video from the car's surround cameras over the last few moments, navigation instructions (for example, "turn left ahead") and the car's own recent motion.
The brain: a 2-billion-parameter reasoning model built on the same Nemotron-based backbone as NVIDIA's Cosmos 3 Edge, sized to fit inside a car's computer.
Outputs: a short explanation of what it sees and why it's acting, which NVIDIA calls Chain-of-Causation reasoning, plus the trajectory the car should follow, drawn by a separate "action expert."
Alpamayo 2 Edge turns camera video, navigation instructions and the car's recent motion into a driving path and, when needed, an explanation. Image: NVIDIA
NVIDIA used a few tricks to shrink the model without losing its smarts. Alpamayo 2 Edge learned its reasoning from its bigger sibling, Alpamayo 2 Super, which generated the explanations it was trained on. A small "draft" model guesses ahead of the main model to speed up its reasoning, and NVIDIA's deployment recipe compresses the model's math to run faster on DRIVE AGX Thor.
The model was also fine-tuned in simulation using NVIDIA's AlpaGym, where it learned from the consequences of its own driving. That step improved its simulated driving performance by 48.9% over its best version without it, according to NVIDIA.
How it compares
NVIDIA pitted Alpamayo 2 Edge against Qwen-Drive-1.0, a model twice its size at 4 billion parameters. In NVIDIA's own benchmarks:
Simulated driving: Alpamayo 2 Edge went 15.15 km (9.4 miles) between at-fault incidents in NVIDIA's AlpaSim simulator, against 0.25 km (0.16 miles) for Qwen-Drive-1.0, roughly 60 times farther.
Path prediction: Alpamayo 2 Edge's planned paths were off from the real driven path by 0.53 m (1.7 ft) on average, compared with 1.68 m (5.5 ft) for Qwen-Drive-1.0.
Scene understanding and reasoning: Alpamayo 2 Edge outscored Qwen-Drive-1.0 in virtually all of NVIDIA's comparisons.
Alpamayo 2 Edge went about 60 times farther between at-fault incidents than Qwen-Drive-1.0 in NVIDIA's simulator. Image: NVIDIA
Note that these results come from NVIDIA's own testing, mostly in simulation, rather than independent road testing.
By the way, it sure looks to us like you are an EV geek like ourselves.
Which means you belong with us: join 14,700+ reading our newsletters: subscribe
What early testers are saying
A handful of companies got early access to Alpamayo 2 Edge, and NVIDIA shared their results:
Niulinx said the model cut its displacement error by 13.6% and its sideways prediction error by 46.1% compared to Alpamayo 2 Super, without any training on Niulinx's own data. Fine-tuning on its own driving data cut displacement error by another 44.2%, and the company went from receiving the model to testing it on a real vehicle within a few days.
Embedl said it got Alpamayo 2 Edge running at 10 Hz, or 10 plans per second, on DRIVE AGX Thor in less than 48 hours.
Alpamayo 2 Edge spots a vehicle cutting in ahead and decides to yield. Image: NVIDIA
UK autonomy company PORT tested the model on recreations of its own London street recordings and saw a clear jump over the previous version:
Within days of getting early access to Alpamayo 2 Edge, we had it running closed-loop in NuRec reconstructions of our own London street recordings. In our early simulation tests, Edge kept its lane in more than twice as many runs as Alpamayo 1.5, recovered from lane offsets about 40% more often, and out of the box planned roughly twice as fast as our hand-tuned Alpamayo 1.5 server on the same GPU.
AlpaJudge grades the AI's reasoning
Shrinking a reasoning model risks making its explanations worse, so NVIDIA also released AlpaJudge, a separate AI that scores how well a driving explanation matches what's actually happening in the scene. AlpaJudge was calibrated against human ratings and comes in two versions: one that compares an explanation against a reference answer, and one that checks it directly against camera frames and the car's maneuvers.
NVIDIA used AlpaJudge in two ways while building Alpamayo 2 Edge:
Picking training data: examples the model already explains well teach it little, while very low scorers can be bad labels. By focusing on middle-scoring clips, NVIDIA raised their share from 27% of the raw data to 81% of the curated training set.
Rewarding good reasoning: used as a reward during training, AlpaJudge improved reasoning quality by 10.2% by its own scoring and by 11.4% as judged by humans, in tests on an Alpamayo 2 Super-based model.
Alpamayo 2 Edge yields to a pedestrian crossing the lane. Image: NVIDIA
Context:
Alpamayo is already finding takers in the autonomy space. Bolt picked NVIDIA Alpamayo to power its European robotaxis, starting with Lucid vehicles in Madrid in 2027, while Einride adopted Alpamayo for its autonomous trucks and aims to be the first trucking company to deploy it on NVIDIA's Hyperion platform.
Alpamayo 2 Edge is the compact member of that family, built to run inside the vehicle rather than in the data center. NVIDIA has published the model and AlpaJudge on Hugging Face, along with a step-by-step recipe on its Alpamayo Developer Hub for getting the model running in real time on DRIVE AGX Thor.
Source: NVIDIA Developer, NVIDIA on LinkedIn
DIG DEEPER: For more robotaxi and autonomy coverage, check out our Robotaxi Report. If you have not already, join 14,700+ EV geeks getting our newsletter: subscribe.
