NVIDIA opens Alpamayo 2 Super for commercial AV model development
NVIDIA is opening Alpamayo 2 Super for commercial model development and redistribution, but has disclosed no customers or deployments.
By Ryan Merket · Published
Why it matters
Commercial licensing gives AV teams broader rights to modify and redistribute NVIDIA's driving model, while the lack of disclosed customers, deployments and independent road validation leaves its production value unproven.
Marco Pavone's autonomous-vehicle research group at NVIDIA is moving Alpamayo 2 Super into commercial development after NVIDIA opened the model for fine-tuning, modification and redistribution on August 4, 2026.
NVIDIA identifies Pavone as its senior director of autonomous-vehicle research and an associate professor at Stanford University. NVIDIA's technical guide says deployed driving policies affect their environment through braking, steering and navigation decisions, allowing errors to compound over time.
That problem has shaped NVIDIA's Alpamayo work. A technical guide describes a development workflow that combines driving models with simulation and reinforcement learning. The aim is to train models against the consequences of their decisions and give developers tools for inspecting why a model selected a maneuver.
The project extends the open-model strategy promoted by NVIDIA co-founder and CEO Jensen Huang. Huang worked at AMD and LSI Logic before starting NVIDIA in 1993 with Chris Malachowsky and Curtis Priem. NVIDIA is a publicly traded company based in Santa Clara, California, and Alpamayo is funded inside the corporation rather than through a separate startup, investor group or financing round, according to its 2026 annual report.
A license change opens the production path
Alpamayo 2 Super is a 32-billion-parameter vision-language-action model for autonomous-driving development, according to NVIDIA's release and model card.
The model is built on Cosmos 3 Super Reasoner and post-trained with reinforcement learning, according to NVIDIA's release. It accepts camera, text and recent ego-motion inputs and is intended for cloud development rather than installation as an untouched, production-ready driving system.
AV developers can use Alpamayo 2 Super as a teacher model to generate trajectories, reasoning traces, labels and training outputs. Those capabilities can then be distilled into smaller models designed for real-time inference inside a vehicle. This cloud-to-car workflow gives automakers and autonomy developers a starting point without requiring them to train every foundation capability from scratch.
Earlier Alpamayo releases carried research-oriented restrictions. NVIDIA has now placed Alpamayo 2 Super under OpenMDW-1.1, a Linux Foundation license, and says it is applying the license across the Alpamayo family. The license terms permit commercial use, modification and distribution of model materials. Redistributors must retain the license and applicable notices, and specified patent or copyright litigation over the materials can terminate certain grants.
NVIDIA says the license lets developers adapt Alpamayo to their data, driving policies and deployment strategies while retaining control of their data and infrastructure and owning the value created through specialized models and accumulated know-how. It also gives suppliers a legal path to package derivative models into commercial products.
Commercial permission does not establish automotive readiness. NVIDIA's model card says developers must conduct use-case-specific testing and system-level validation before deployment. Alpamayo 2 Super remains a cloud development model and a source for distilling capabilities into smaller in-vehicle systems. NVIDIA has disclosed no pricing, revenue, paying customers or commercial deployments for Alpamayo 2 Super, and it has not named an automaker, truckmaker, supplier or AV operator using the commercially licensed release.
One model handles five AV development tasks
Alpamayo 2 Super processes synchronized feeds from cameras around a vehicle, along with text and recent ego-motion history. It can return a planned trajectory, a chain-of-causation explanation, a high-level action such as yielding or stopping, automated reasoning labels, and visual answers grounded in particular image regions.
These outputs address several AV workflows. A developer can inspect the model's proposed path, examine its stated reasoning, use it to label fleet video and compare the output with a smaller driving policy.
Difficult cases such as cut-ins, complex intersections, pedestrian interactions and adverse weather form the technical case for a reasoning model. Conventional systems can learn frequent patterns from large driving datasets. Rare interactions involving several road users are harder because the vehicle must infer intent, anticipate consequences and select a safe action from limited examples.
NVIDIA also connects Alpamayo's reasoning traces to its Halos safety-validation tools and work aligned with ISO/PAS 8800, according to the company's product materials. That integration may help teams organize validation evidence, but it does not mean Alpamayo 2 Super is certified or approved for public-road deployment.
NVIDIA's benchmarks remain NVIDIA's benchmarks
According to the Alpamayo 2 Super model card, NVIDIA measured a 79.2 score on LingoQA's Lingo-Judge metric and ranked the model first among nearly 40 systems it evaluated. NVIDIA reported that Alpamayo 2 Super exceeded Qwen2.5-VL 72B by 17 points, Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points.
These results show how NVIDIA evaluates the model inside its own development stack. They do not demonstrate safe operation on public roads, regulatory approval or performance inside a complete redundant vehicle system. Independent replication will matter because the model, reported comparisons and much of the evaluation infrastructure come from NVIDIA.
A 2026 research paper found that reasoning-enabled autonomous-driving models, including Alpamayo, could be disrupted by manipulated textual inputs. The researchers called for stronger defenses and more rigorous evaluation of the connection between model reasoning and vehicle trajectories. Open weights make that kind of outside testing easier, while commercial licensing widens the group of developers who can examine and modify the system.
Alpamayo competes for developers, not road miles
Alpamayo enters a market where several rivals are building complete autonomy products. Waymo provides the clearest deployment comparison through its proprietary, vertically integrated commercial robotaxi service. Wayve is developing an end-to-end AI Driver for use across vehicle platforms and says its announced consumer-vehicle deployment begins in 2027. Waabi is initially applying its end-to-end Physical AI system to autonomous trucking, with robotaxis planned as a later market. Applied Intuition sells embedded autonomy software alongside simulation and fleet-management products.
Pavone's group is offering a development model and associated tools rather than operating a fleet or selling a complete certified driving system. Its competitive argument rests on open weights, commercial redistribution rights and access to NVIDIA's training and simulation stack. With no disclosed customers or deployments, the public evidence around the release is limited to NVIDIA's benchmark reports and technical materials.
Pavone is assembling an AV development stack
NVIDIA introduced the broader Alpamayo family in late 2025, before the August 2026 commercial-licensing release. A technical overview describes Alpamayo 2 Super alongside AlpaSim, a closed-loop simulator, and AlpaGym, a reinforcement-learning framework that feeds simulated driving consequences back into model training.
The currently published closed-loop training guide demonstrates AlpaGym with Alpamayo 1.5, a 10-billion-parameter predecessor. NVIDIA says developers can use their own AV models, while broader workflows remain under development. The documentation therefore trails the commercial release of Alpamayo 2 Super's larger model.
The surrounding tools give AV teams a path through data processing, model training, simulation, distillation and in-vehicle inference. Open weights reduce the legal friction of entering that stack, while specialized models distilled from Alpamayo can create demand for NVIDIA compute in data centers and vehicles. Developers can inspect parts of the supporting work through the Alpamayo GitHub repository.
The August 4 license change turns research adoption into a commercial test. Pavone's group has provided the model, training workflow and simulation tools, and automakers and AV developers can now build proprietary systems on top without negotiating separate model rights. They still have to establish that reasoning traces, benchmark scores and simulated driving gains hold up inside a safety-critical vehicle.