NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use

For developers of robotaxis and autonomous vehicles (AVs), the most daunting challenges are rarely found in routine, predictable driving. Instead, the industry’s greatest hurdle remains the "long-tail" of rare, complex, and unpredictable scenarios that are notoriously difficult to anticipate during training. Successfully navigating these edge cases requires more than simple object detection or basic motion prediction; it demands that an AV truly understands its environment, reasons through cause and effect, executes the appropriate maneuver, and translates that decision into a path that is both safe and comfortable—all in real-time, while remaining transparent enough for developers to validate and trust.

To address these needs, NVIDIA has officially released Alpamayo 2 Super for commercial use. As the latest addition to the Alpamayo family—already the most widely adopted open reasoning model suite for autonomous driving on Hugging Face—this foundation model is designed to support a vast array of AV-critical tasks within a single, unified architecture. Built upon the NVIDIA Cosmos 3 Super Reasoner and refined through advanced reinforcement learning, Alpamayo 2 Super represents a significant leap forward for the AV ecosystem, combining open commercial licensing with state-of-the-art multitask capabilities.

A New Era of Open Licensing for Production AVs

A defining feature of this release is its accessibility. Alpamayo 2 Super is now available on Hugging Face under the OpenMDW-1.1 license, the Linux Foundation’s permissive framework for open AI model distribution. This license explicitly permits fine-tuning, the creation of derivative models, and commercial redistribution. This move empowers automakers, truck manufacturers, and specialized AV suppliers to tailor the Alpamayo architecture to their unique datasets, proprietary driving policies, and specific deployment strategies.

"Openness allows AV researchers and companies to maintain full control over their data and infrastructure, while retaining the value they generate through specialized models and proprietary know-how. Such control is non-negotiable for workflows involving commercial fleets and rigorous safety standards."

By applying the OpenMDW license across the entire Alpamayo family, NVIDIA has created a seamless bridge between research and production. Developers no longer need to navigate complex permission structures to deploy these models commercially. Furthermore, the open-weight nature of the project makes the transition from development to deployment economically viable, allowing teams to leverage advanced reasoning without the prohibitive costs of training foundation models from scratch.

The Cloud-to-Car Workflow

The Alpamayo family is engineered to support a "cloud-to-car" development cycle. Alpamayo 2 Super provides frontier-scale reasoning in cloud-based environments, where developers can generate high-fidelity reasoning traces, synthetic training data, and teacher outputs for model distillation.

While Alpamayo 2 Super offers the highest performance for complex reasoning, the family also includes Alpamayo 1.5 and Alpamayo 1, which serve as cost-efficient alternatives for specific cloud tasks and model distillation. Once distilled, these models can be optimized for efficient, real-time inference directly on the vehicle’s edge hardware. This tiered approach provides a sustainable path for scaling safe autonomy across massive commercial fleets.

Benchmark-Leading Reasoning at Scale

In terms of raw performance, Alpamayo 2 Super is currently unrivaled in its class. It holds the top position on the LingoQA autonomous driving reasoning benchmark, outperforming nearly 40 other models. According to NVIDIA’s internal testing using the Lingo-Judge metric, the model significantly outperformed industry heavyweights:

• 17.0 points ahead of Qwen2.5-VL 72B.
• 15.1 points ahead of Gemini 2.5 Pro.
• 23.2 points ahead of GPT-4o.

With three times the scale of the 10-billion-parameter Alpamayo 1.5 and 1 models, Alpamayo 2 Super excels at generalizing from sparse data—a critical requirement for handling multi-agent interactions where traditional systems often fail. By processing full-surround camera coverage, the model maintains a 360-degree context, enabling a deeper understanding of high-risk maneuvers like unprotected turns, merges, and complex, multi-lane intersections.

A Multitask Powerhouse for Autonomous Driving

For every driving scenario, Alpamayo 2 Super generates five tightly integrated outputs that provide a window into the "mind" of the vehicle:

1. Trajectory Planning: A precise path for the vehicle to follow. 2. Chain-of-Causation (CoC) Trace: A logical explanation of the reasoning behind every decision. 3. Meta-Action: A clear classification of intent, such as "yield," "stop," or "lane change." 4. Reasoning Auto-labels: Automated annotations that generate CoC data for future training and validation. 5. Visual Question Answering (VQA): 2D visual grounding that links the model’s reasoning to specific regions within the camera feed.

These outputs allow developers to correlate what the model observed with the action it chose, making the decision-making process easier to critique and validate. Furthermore, the CoC traces integrate directly with NVIDIA Halos safety-validation workflows, supporting AI safety alignment with ISO/PAS 8800 requirements.

Beyond planning, the model acts as a powerful auto-labeler. By transforming raw, unlabelled driving clips into rich, annotated training data, Alpamayo 2 Super can compress annotation cycles from months down to mere days.

Building the Future of Autonomy

Alpamayo 2 Super is just one component of a broader, open ecosystem for AV development. It works in tandem with:

• NVIDIA AlpaSim: For closed-loop simulation.
• NVIDIA AlpaGym: For high-throughput reinforcement learning.
• NVIDIA Physical AI Open Datasets: Providing the raw material for training and testing.

With over 500,000 downloads already recorded on Hugging Face, the Alpamayo family has solidified its status as the industry standard for open reasoning in autonomous driving. By providing the tools to reason, validate, and scale, NVIDIA is clearing the path for the next generation of safe, reliable, and commercially viable robotaxis.