Nvidia has launched Alpamayo 2 Super, a new open reasoning model for autonomous vehicles that the chip giant says can analyze complicated driving situations, reason through what is happening around a vehicle, and decide what it should do next.
The model is available for commercial use under the OpenMDW-1.1 license, giving automakers, autonomous driving companies and developers permission to inspect, fine-tune, modify and deploy it. Nvidia is positioning Alpamayo 2 Super as a foundation model for robotaxis, autonomous trucks, shuttles, delivery vehicles, tractors and other mobile robots.
That commercial license may prove just as significant as the model itself.
Autonomous driving companies have spent years building systems that can recognize pedestrians, traffic lights, lanes and surrounding vehicles. The harder problem sits at the edges: the unusual situation that appears rarely in training data but still requires the vehicle to make the right decision in seconds.
Nvidia wants Alpamayo 2 Super to reason through those moments.
Announcing the release on X, Nvidia CEO Jensen Huang described the idea in unusually simple terms: the model should “think before it acts.”
“Today, we’re launching Alpamayo 2 Super, our frontier open reasoning model for autonomous vehicles. Beyond seeing, Alpamayo understands and reasons through the complex world – thinks before it acts.”
Huang added:
“It’s a powerful backbone for robotaxis, trucks, shuttles, delivery vans, tractors and the long tail of mobile robots—billions of autonomous machines someday. We’re releasing it for commercial use under OpenMDW-1.1 so teams can inspect it, fine-tune it and deploy it—open models advance safety and security. The next wave of AI is robotics—and it starts with autonomous vehicles. Great work, Alpamayo team!”
Today, we’re launching Alpamayo 2 Super, our frontier open reasoning model for autonomous vehicles.
Beyond seeing, Alpamayo understands and reasons through the complex world – thinks before it acts.
It’s a powerful backbone for robotaxis, trucks, shuttles, delivery vans,… pic.twitter.com/K6jS0XDR4b
— Jensen Huang (@JensenHuang) August 4, 2026
The release pushes Nvidia deeper into one of AI’s most difficult physical applications. A chatbot can generate a bad answer and try again. A vehicle traveling through an intersection does not have that luxury.
From seeing the road to reasoning about it
Modern autonomous driving systems are already good at many routine perception tasks. Cameras and other sensors identify cars, cyclists, pedestrians, traffic signals, road markings, and obstacles. Prediction systems estimate where surrounding objects may move. Planning software chooses a path.
The problems become harder when those signals do not produce an obvious answer.
A pedestrian may hesitate near a crosswalk. A vehicle may partially enter an intersection before stopping. Two drivers may attempt to merge into the same lane. Construction may create a road configuration that barely resembles the map or training examples.
These long-tail scenarios have been one of the stubborn problems facing autonomous driving developers.
NVIDIA Alpamayo 2 Super
Nvidia says Alpamayo 2 Super approaches the problem through reasoning. Rather than producing a driving trajectory alone, the model can generate information explaining the chain of causation behind its decision.
That distinction matters for autonomous vehicle development. Engineers need more than a system that produces the correct answer most of the time. They need ways to examine why it produced that answer, test the reasoning against difficult scenarios, and identify failure modes before software reaches public roads.
Alpamayo 2 Super processes full-surround camera coverage from the front, sides and rear of a vehicle. Nvidia says the 360-degree view gives the model more context for situations such as lane changes, merges, unprotected turns and crowded intersections.
Alpamayo 2 Super can explain what it plans to do
For a driving scenario, Alpamayo 2 Super can generate a planned vehicle trajectory together with a chain-of-causation trace explaining the reasoning behind that decision.
The model can identify a meta-action such as yielding, stopping or changing lanes. It can create reasoning labels for training and validation data, and answer visual questions tied to specific areas of camera images.
Taken together, those capabilities give developers a way to connect three questions that are often separated inside an autonomous driving stack: What did the vehicle see? What did the model infer from it? What action did it choose?
That could become particularly useful during safety validation.
If an autonomous vehicle chooses to yield at an intersection, for example, engineers can inspect the reasoning associated with that decision rather than treating the model’s output as an opaque command. Nvidia says its chain-of-causation traces can integrate with the company’s Halos safety-validation workflows and support AI safety work aligned with ISO/PAS 8800 requirements.
There is another use for the model that could matter before any vehicle reaches the road.
Alpamayo 2 Super can act as an autolabeler for proprietary fleet footage. Nvidia says the model can analyze driving clips, generate chain-of-causation labels and connect its answers to regions within camera images.
Training-data annotation is expensive and time-consuming for AV companies. Nvidia claims this approach can compress annotation cycles that once took months into days.
Nvidia is opening Alpamayo for commercial use
The licensing change separates Alpamayo 2 Super from Nvidia’s earlier releases, which initially focused on research and development.
Alpamayo 2 Super is being released on Hugging Face under OpenMDW-1.1, a permissive license for open AI model distributions from the Linux Foundation. Nvidia says the license permits fine-tuning, derivative models, and commercial redistribution.
The company is applying the license across the Alpamayo model family, creating a path for companies that experimented with the technology during development to move their work into commercial deployments without seeking separate permission from Nvidia.
For automakers and autonomous vehicle companies, open weights could change the economics of that process.
Companies can adapt Alpamayo models using proprietary fleet data and driving policies rather than rebuilding every foundation capability themselves. They can keep those specialized models inside their own infrastructure and retain control over the intellectual property created on top of Nvidia’s foundation models.
That structure fits the economics Nvidia is proposing for autonomous driving AI: large reasoning models running during cloud-based development, followed by smaller specialized models distilled for real-time use inside vehicles.
A cloud-to-car AI strategy
Alpamayo 2 Super is the largest reasoning model in the family.
Nvidia says it has three times the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models. The extra capacity is aimed at improving reasoning across sparse examples, including rare interactions between multiple road users.
Running a model of that scale directly inside millions of vehicles would create obvious computing and cost challenges.
Nvidia’s answer is model distillation.
Developers can use Alpamayo 2 Super in cloud environments to produce reasoning traces, synthetic training data, and teacher outputs. Smaller models can then learn from those outputs and be optimized for lower-latency inference inside production vehicles.
The approach effectively separates the expensive reasoning process used during development from the model that eventually needs to make decisions inside a car.
Alpamayo 1 and Alpamayo 1.5 remain available as less expensive options for development and distillation, giving companies several model sizes depending on their workloads.
For Nvidia, the architecture creates a connection between its AI infrastructure business and the growing physical AI market. Training, simulation and large-scale reasoning can happen in Nvidia-backed data centers, with optimized models eventually running on automotive computing platforms inside vehicles.
Nvidia says Alpamayo beats Gemini and GPT-4o on a driving benchmark
Nvidia is making aggressive performance claims for its newest model.
According to Nvidia’s testing, Alpamayo 2 Super ranks first on LingoQA, a benchmark focused on reasoning for autonomous driving, among nearly 40 models evaluated.
Using the Lingo-Judge metric, Nvidia says Alpamayo 2 Super beat Qwen2.5-VL 72B by 17 points, Google’s Gemini 2.5 Pro by 15.1 points, and OpenAI’s GPT-4o by 23.2 points.
Those results should be read in context. They are Nvidia-reported benchmark results, and strong performance on an autonomous driving benchmark does not by itself establish that a model is ready to safely control vehicles on public roads.
Real-world autonomous driving involves hardware failures, changing weather, unusual road behavior, sensor limitations and interactions that no benchmark can fully reproduce.
Still, the comparison points to Nvidia’s larger argument. General-purpose multimodal models can reason about images and video, but models trained around driving may have an advantage when the questions involve vehicle behavior, road geometry and the intentions of multiple road users.
Nvidia is building more than an autonomous driving model
Alpamayo 2 Super sits inside a larger collection of Nvidia technology aimed at autonomous vehicle development.
AlpaSim provides closed-loop simulation for testing autonomous systems. AlpaGym supports reinforcement learning at scale. Nvidia’s Physical AI Open Datasets provide training and evaluation data, with open training recipes and autolabeling tools filling out more of the development process.
The pieces reveal Nvidia’s broader strategy.
The company does not need to build the robotaxi itself to have a major role in the robotaxi economy.
Nvidia can supply the computing hardware, foundation models, simulation environment, training tools, safety systems and in-vehicle computing platforms used by companies that build autonomous machines.
It is a familiar Nvidia playbook. Rather than betting on a single application, the company builds infrastructure that many competing companies can use.
Alpamayo’s adoption gives Nvidia an early foothold. The model family has passed 500,000 downloads on Hugging Face, according to the company, which says Alpamayo is now the platform’s most-adopted family of open reasoning models for autonomous driving.
The real test starts outside the benchmark
Autonomous vehicles have spent years moving through cycles of enormous expectations followed by reminders of just how difficult driving actually is.
AI reasoning models introduce another approach to the problem, but they do not erase the safety, regulatory, and engineering hurdles facing robotaxis and self-driving vehicles.
What Alpamayo 2 Super changes is access.
Developers can now take Nvidia’s largest autonomous-driving reasoning model, inspect its weights, adapt it to proprietary data, generate training material from it and use the resulting technology commercially.
That creates an experiment far larger than anything Nvidia could run alone.
Automakers, trucking companies, robotaxi developers and robotics startups can now test whether reasoning models really help autonomous machines deal with the strange, ambiguous moments that conventional driving systems struggle to handle.
If they do, the next major leap in autonomous driving may come less from teaching vehicles to see more objects and more from teaching them to make better sense of what they already see.
For Nvidia, that is the bet behind Alpamayo 2 Super: the autonomous vehicle of the future shouldn’t simply recognize the road around it. It should be able to explain why it decided to turn, stop, or wait before it acts.
