NVIDIA has released Alpamayo 2 Super for commercial use, an open autonomous driving reasoning model roughly three times the scale of the company’s 10-billion-parameter Alpamayo 1.5 and Alpamayo 1. The model is distributed on Hugging Face under OpenMDW-1.1, the Linux Foundation’s permissive license for open AI model distributions. It is built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, and NVIDIA says it leads every autonomous driving benchmark the company has evaluated it against.
Highlights
- Released under OpenMDW-1.1, covering fine-tuning, derivative models and commercial redistribution
- Approximately 3x the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1
- Ranked first on LingoQA among nearly 40 models evaluated, per NVIDIA
- Alpamayo family downloads have passed 500,000 on Hugging Face
Open Licensing Across the Full Model Family
Earlier Alpamayo releases were introduced for research and development only. NVIDIA is now applying the OpenMDW license across the entire Alpamayo family, meaning developers can deploy any model in the lineup commercially without seeking additional permissions. The license covers fine-tuning, derivative models and commercial redistribution, which the company says allows AV developers, automakers, truckmakers and suppliers to adapt the models to their own data, driving policies and deployment strategies.
NVIDIA frames that control as essential for workflows built on proprietary fleets and safety validation, where teams need to retain ownership of both their data and the specialized models they derive from it. Open weights also change the cost structure: teams can build on the reasoning capability without retraining foundation behavior from scratch or paying frontier-model rates for every task.
Within the family, the company positions Alpamayo 2 Super as the highest-performing option for cloud-based development — generating reasoning traces, synthetic training data and teacher outputs for distillation — while Alpamayo 1.5 and Alpamayo 1 remain the more cost-efficient choices for the same workflows. Distilled models can then be tuned for real-time inference in production vehicles, forming what NVIDIA describes as a cloud-to-car pipeline.
How Does Alpamayo 2 Super Perform on Driving Benchmarks?
NVIDIA reports that the model ranks first on LingoQA, an autonomous driving reasoning benchmark, among nearly 40 models evaluated. In the company’s own testing using the Lingo-Judge metric, it outperformed:
- Qwen2.5-VL 72B by 17.0 points
- Gemini 2.5 Pro by 15.1 points
- GPT-4o by 23.2 points
The company also states the model ranks first across all autonomous driving benchmarks it has evaluated. The added parameter count, NVIDIA says, helps the model generalize reasoning from sparse examples — the rare, multi-agent interactions that conventional stacks handle poorly. Perception spans full-surround camera coverage, fusing front, side and rear views into a 360-degree picture for lane changes, merges, unprotected turns and complex intersections.
Five Coupled Outputs Per Driving Situation
For robotaxis and other autonomous vehicles, the hardest problems are the long-tail events that resist anticipation. NVIDIA’s argument is that handling them requires more than object detection and motion prediction — it requires reasoning a developer can inspect.
To that end, Alpamayo 2 Super produces five tightly coupled outputs for each driving situation:
- Trajectory describing the vehicle’s planned path
- Chain-of-causation (CoC) trace explaining the reasoning behind the decision
- Meta-action capturing intent, such as yield, lane change or stop
- Reasoning auto-labels generating CoC annotations for training and validation data
- Visual question answering with 2D visual grounding, linking answers to specific regions in camera images
Together, these let developers tie what the model observed to the action it selected.
Autolabeling and Safety Workflow Integration
CoC traces integrate with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800 requirements. Deployed as an autolabeler on proprietary fleet data, the model generates CoC labels and performs grounded visual question answering — work NVIDIA says compresses annotation cycles from months to days.
Beyond planning and auto-labeling, the model supports scene understanding, model critiquing and knowledge distillation, allowing a single foundation model to serve multiple stages of the development stack.
The Surrounding Toolset
Alpamayo 2 Super sits within a broader set of NVIDIA open models, frameworks and datasets for AV development. That collection includes NVIDIA AlpaSim for closed-loop simulation, NVIDIA AlpaGym for high-throughput reinforcement learning, NVIDIA Physical AI Open Datasets for training and testing data, and open training recipes paired with an autolabeling pipeline. Additional tools for autonomous driving are available through NVIDIA’s developer portal.
Sign up for our weekly emails to stay up to date on the latest news!
