NVIDIA Alpamayo 2 Super: A Leading Open Model for Autonomous Vehicle Development

For autonomous vehicles (AVs) and robotaxis, addressing complex, infrequent scenarios poses a significant challenge. While everyday driving conditions are manageable through standard object detection and motion prediction, the unpredictable long-tail events require a more sophisticated understanding of context, causality, and decision-making.

The introduction of the NVIDIA Alpamayo 2 Super model marks a notable advancement in this arena. This model is now available for commercial use and is a part of the Alpamayo series, which has become the most widely adopted open reasoning models for automated driving on Hugging Face. It integrates a range of AV-related capabilities within a single framework.

Constructed on the NVIDIA Cosmos 3 Super Reasoner and enhanced with reinforcement learning techniques, Alpamayo 2 Super pushes the limits of the AV ecosystem on two primary fronts: it offers open commercial licenses and delivers superior multitasking abilities tailored for autonomous driving applications.

Open Licensing Enhancements for AV Production

The Alpamayo 2 Super is distributed under the OpenMDW‑1.1 license, a permissive agreement governed by the Linux Foundation that facilitates open AI model distributions. This licensing entails provisions for fine-tuning, creating derivative models, and commercial redistribution, allowing AV developers and manufacturers to tailor the model to their specific datasets, operational methodologies, and deployment scenarios.

This level of openness empowers companies to retain control over their proprietary data and infrastructures, while also ensuring they benefit from the innovations born from custom adaptations and knowledge accumulation. Such control is instrumental in maintaining safety and efficiency in environments laden with proprietary AV fleets.

Previous versions of Alpamayo were primarily research-focused, but with the updated OpenMDW framework applied across the board, developers can transit directly from adapting the models to deploying them commercially without additional permissions. This seamless transition enhances agility in developing and scaling AV solutions.

Cost-effective access is facilitated by open weights, enabling teams to leverage advanced reasoning capabilities without the exorbitant costs typically associated with cutting-edge model development. Consequently, organizations can appropriate daily operational costs to match specific tasks without requiring unnecessary retraining processes for every fundamental component.

Within cloud-based development environments, Alpamayo 2 Super provides frontier-scale reasoning, allowing developers to create high-quality reasoning traces, synthetic datasets, and teacher outputs for model distillation. Notably, it outperforms its predecessors, Alpamayo 1.5 and Alpamayo 1, by providing enhanced reasoning and driving performance for multimodal AV development.

Leading Benchmark Performances in Autonomous Driving

In recent evaluations, Alpamayo 2 Super emerged as a top-performer on LingoQA, a reasoning benchmark specifically designed for autonomous driving, outperforming nearly forty models assessed. It surpassed models like Qwen2.5‑VL 72B, Gemini 2.5 Pro, and GPT‑4o by 17.0 points, 15.1 points, and 23.2 points, respectively, showcasing its capacity for advanced reasoning in driving-centric scenarios.

Notably, it boasts a parameter capacity three times that of its predecessor, the 10-billion-parameter Alpamayo 1.5. This enhancement augments the model's ability to generalize reasoning from limited examples, a vital capability when dealing with infrequent, multi-agent interactions that often challenge existing systems.

Equipped with comprehensive sensor data from full-surround camera views, Alpamayo 2 Super excels in situations that entail lane changes, merges, and complex intersections, where understanding context is critical for ensuring safety.

Multitasking Capabilities for Development Flexibility

Alpamayo 2 Super is engineered to generate five integrated outputs for each driving scenario:

• A trajectory detailing the vehicle's planned path.

• A chain-of-causation (CoC) trace elucidating the reasoning behind decisions made.

• A meta-action that encapsulates the model's intent (e.g., yielding, lane changes).

• Reasoning auto-labels that create CoC annotations for training and validation of data.

• Visual question answering responses that correspond with specific areas within camera images.

These outputs provide transparency in decision-making processes, allowing developers to understand how observed data correlates with chosen actions, facilitating easier validation and critique. CoC traces seamlessly integrate into NVIDIA Halos safety-validation workflows, reinforcing AI safety aligned with ISO/PAS 8800 standards.

Additionally, Alpamayo 2 Super serves as an autolabeler, useful for generating CoC labels and conducting visual question answering with 2D grounding on proprietary dataset inputs. By connecting reasoning to visual data, the model can effectively transform raw driving footage into training materials, significantly reducing annotation timelines from months to just days.

A Comprehensive Ecosystem for Reasoning-based AV Development

Part of a more extensive family of open models, frameworks, and datasets, Alpamayo 2 Super is poised to revolutionize AV development. Supplementary tools include:

• NVIDIA AlpaSim: Provides closed-loop simulations for testing AV behaviors.

• NVIDIA AlpaGym: Supports high-throughput reinforcement learning for model optimization.

• NVIDIA Physical AI Open Datasets: Delivers essential training and testing data for enhanced model performance.

• Open training recipes and an autolabeling pipeline: These resources accelerate the development, training, and validation phases.

With over 500,000 downloads on Hugging Face, Alpamayo has solidly established itself as the leading open reasoning model for AVs on the platform. Developers looking to explore the model's capabilities can obtain NVIDIA Alpamayo 2 Super on Hugging Face to evaluate its performance and contribute to the evolution of robotaxis and autonomous vehicles.