NVIDIA has released Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, under an open commercial license designed to tackle the long-tail events that conventional detection-and-prediction stacks handle poorly. The model, available on Hugging Face, pairs a 32B VLM backbone built on NVIDIA Cosmos 3 Super Reasoner with a 2.3B diffusion-based action decoder, and from one pass over full-surround camera video it emits a planned trajectory, a causal explanation of that trajectory, and a meta-action. This release signals a significant shift in how autonomous driving systems are being architected — moving away from fragmented pipelines toward unified models that can reason about what they see and explain why they act.
Open Commercial Availability Under OpenMDW-1.1
Alpamayo 2 Super is deployable for commercial use immediately. The weights are released under OpenMDW-1.1, the Linux Foundation’s permissive license for open model distributions, and the source code is Apache 2.0. The license covers fine-tuning, derivative models, and commercial redistribution without requiring additional permission from NVIDIA. This is a notable departure from earlier releases in the Alpamayo family, which were introduced for research and development purposes. NVIDIA is now applying OpenMDW across the entire Alpamayo family, meaning that earlier releases once restricted to R&D are now commercially deployable. For engineering teams evaluating autonomous driving stacks, this removes a significant legal barrier and accelerates the path from evaluation to production.
Model Architecture and Input Specifications
Understanding what goes into Alpamayo 2 Super and what comes out is essential for any team considering integration. The model accepts multi-camera RGB video, text, and egomotion history with timestamps. The validated public notebook profiles use six cameras and four historical frames per camera, while the measured inference profile used seven cameras. Egomotion data is structured as 3D translation plus a 3×3 rotation matrix across multiple timesteps, providing the model with a precise understanding of vehicle movement over time.
The trajectory API returns 64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals. Each waypoint carries ego-frame XYZ coordinates and a 3×3 rotation matrix, giving downstream systems the full pose information needed for path planning and control. The combination of visual input, egomotion history, and structured trajectory output makes the model suitable for direct integration into existing autonomous driving software stacks.
Training Data Scale and Composition
The scale of training data behind Alpamayo 2 Super is substantial. The model was trained on roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations. This dataset includes approximately 3.7 million Chain-of-Causation (CoC) traces — structured, causally linked explanations of driving decisions that teach the model not just what to do, but why. Image training data exceeds one billion images, providing the visual foundation needed for robust perception across diverse environments, lighting conditions, and weather scenarios.
What is particularly interesting from a technical standpoint is the CoC data structure. These traces are not simple labels; they are causally linked explanations that connect observed events to driving decisions. For example, a CoC trace might explain that a vehicle is slowing because a pedestrian is approaching a crosswalk, and that the ego vehicle is planning a stop to yield. This causal reasoning capability is what differentiates Alpamayo 2 Super from earlier vision-language models and from traditional autonomous driving pipelines that separate perception, prediction, and planning into distinct modules.
Benchmark Performance: Numbers That Matter for Planning Work
On LingoQA, a benchmark for evaluating language models on driving-related visual question answering, Alpamayo 2 Super records a Lingo-Judge score of 79.2 and ranks first among nearly 40 models evaluated. In NVIDIA’s testing, it outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. These margins are substantial and indicate that the model’s reasoning capabilities over driving scenes are not merely competitive but meaningfully ahead of general-purpose models.
Two additional metrics matter specifically for planning work. In closed-loop evaluation with AlpaSim on 910 scenarios from the PhysicalAI-AV-NuRec dataset, the model achieves an AlpaSim score of 1.50 ± 0.13. Closed-loop evaluation is particularly important because it measures how the model performs when its outputs feed back into a simulated environment, creating a more realistic assessment of real-world behavior. In open-loop evaluation on 937 challenging samples from the PhysicalAI-AV dataset, the model achieves a minADE₆ at 6.4s of 0.911 meters. minADE stands for minimum Average Displacement Error, and a value below one meter over a 6.4-second horizon is a strong result that suggests the trajectory predictions are spatially accurate even over longer prediction windows.
How These Benchmarks Translate to Real-World Deployment
For engineering teams evaluating this model, it is useful to separate headline metrics from operational considerations. The LingoQA score indicates strong scene understanding and reasoning capabilities, which directly support explainability and safety validation. The AlpaSim closed-loop score suggests the model can operate reliably within a simulated environment, handling scenario variations without catastrophic failures. The minADE₆ metric confirms that trajectory outputs are accurate enough for downstream planning modules to use without extensive post-processing.
The measured inference profile used one NVIDIA H100 80GB GPU at 72,115 MiB peak memory in BF16 precision. This is a cloud-scale model, not an edge-scale model. Teams planning in-car deployment will need to distill the model to smaller architectures or explore quantization and pruning techniques. NVIDIA’s documentation supports this path, and the open license does not restrict derivative models, making distillation workflows commercially viable.
Five Outputs From One Model: A Unified Approach to Driving Decisions
For each driving situation, Alpamayo 2 Super produces five distinct outputs from a single inference pass. The first is a trajectory, which is the planned path of the vehicle over the next 6.4 seconds. The second is a CoC trace, which explains the reasoning behind that trajectory in a structured, causally linked format. The third is a meta-action — a compact intent label such as yield, lane change, or stop that summarizes the high-level behavior. The fourth output is reasoning auto-labels, which are CoC annotations generated automatically over proprietary fleet data. The fifth is visual question answering with 2D grounding, enabling the model to answer questions about a scene while referencing specific regions in the camera images.
This combination of outputs is operationally valuable. Developers can tie what the model observed to the action it chose, which is critical for debugging, safety validation, and regulatory compliance. The CoC traces integrate with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800, a standard for safety of AI systems in road vehicles. This means the model is not just producing outputs — it is producing audit-ready explanations that can be used in formal safety cases.
Autolabeling: Compressing Annotation Cycles From Months to Days
One of the most immediately practical applications is the use of Alpamayo 2 Super as an autolabeler on proprietary fleet data. NVIDIA states that the model compresses annotation cycles from months to days. For companies operating large vehicle fleets, this capability has direct cost implications. Manual annotation of driving data is expensive, time-consuming, and prone to inconsistency. An automated system that can generate CoC traces and trajectory labels at scale reduces both cost and time-to-market for model updates.
This is particularly relevant for teams that already collect large amounts of fleet data but lack the resources to annotate it comprehensively. Alpamayo 2 Super can generate structured training data that is then used to fine-tune smaller, more efficient models for specific operational domains. This workflow — large open model generates labels, smaller proprietary model learns from those labels — is becoming a standard pattern in autonomous driving development.
What Is the Chain-of-Causation Architecture and Why It Matters
The Chain-of-Causation architecture is a structured approach to explaining driving decisions. Unlike simple text descriptions, CoC traces establish explicit causal links between observed events, model reasoning, and chosen actions. A trace might describe a sequence: the ego vehicle observes a lead vehicle braking, reasons that the lead vehicle is likely slowing for a red light, and decides to maintain a safe following distance with a planned deceleration. Each step in this chain is explicitly connected, creating a complete causal narrative that humans can review.
This architecture matters because it directly addresses a core challenge in autonomous driving: trust. Regulators, safety auditors, and the general public need to understand why an autonomous vehicle makes specific decisions, particularly in edge cases. CoC traces provide the transparency needed for certification and for post-incident analysis. When paired with the 2D grounding capability, engineers can visually inspect exactly which regions of the camera images the model attended to when making decisions.
Reference Architecture for Integration Workflows
For teams planning to integrate Alpamayo 2 Super into their workflows, the validated public notebooks provide a reference architecture. The six-camera configuration with four historical frames per camera represents a balanced trade-off between computational cost and temporal context. The inclusion of egomotion history is important — it gives the model the ability to reason about vehicle dynamics rather than treating each frame as an independent snapshot.
Running this model requires substantial compute resources. At 34B parameters in BF16 precision, teams will need multiple high-end GPUs for inference, and fine-tuning will require even more resources. NVIDIA’s documentation addresses this through two distinct paths. The first is distillation to smaller architectures for in-car deployment. The second is using the model as a label generator for proprietary data pipelines, where the cloud-scale cost can be amortized across many downstream uses.
Hardware Requirements and Deployment Considerations
The model’s peak memory footprint of 72,115 MiB on a single H100 80GB GPU indicates that teams should plan for data-center-scale inference rather than edge deployment. In practice this means the model is best used for offline processing, cloud-based fleet data analysis, and generating training labels. For real-time in-car inference, distillation to a smaller model is the recommended path. The open license permits this, and NVIDIA’s ecosystem provides tooling for model compression and quantization.
Teams should also note that the model’s performance in closed-loop evaluation was conducted on a specific dataset (PhysicalAI-AV-NuRec) with specific scenarios (910 in total). Real-world performance may vary based on geography, traffic patterns, and vehicle platform. Thorough validation on proprietary data is essential before deployment, and the CoC trace outputs provide the audit trail needed for that validation process.
Market Context: The Shift From Modular Pipelines to Unified VLA Models
The release of Alpamayo 2 Super is part of a broader industry trend toward unified vision-language-action models for robotics and autonomous driving. Traditional autonomous driving stacks separate perception, prediction, planning, and control into distinct modules, each with its own models and interfaces. This modular approach has been successful in many ways, but it struggles with long-tail events — rare situations where the interaction between modules becomes difficult to manage, and where emergent behaviors can be unpredictable.
VLA models take a different approach. They treat driving as a single end-to-end task, where raw sensor input is mapped directly to action. The language component adds a new dimension: the ability to explain actions in human-readable form. This is not just about user interfaces or regulatory compliance. Language-enabled reasoning allows the model to generalize from known situations to novel ones by leveraging world knowledge embedded in the language model.
The commercial release under OpenMDW-1.1 is strategically significant because it lowers the barrier for adoption across the industry. Competing models remain largely proprietary, and engineering teams are often locked into vendor-specific ecosystems. An open model with permissive licensing gives teams the freedom to experiment, fine-tune, and deploy without contractual constraints. This could accelerate adoption of VLA architectures across the autonomous driving industry.
Comparison With Previous Alpamayo Releases and the Evolution Within the Family
The Alpamayo family has evolved through several iterations, each addressing limitations of its predecessor. Earlier releases were introduced for research and development, with restrictions that limited commercial applications. The application of OpenMDW-1.1 across the entire family represents a philosophical shift in NVIDIA’s approach to the autonomous driving market — moving from a protective stance toward an aggressive adoption strategy. By making the family commercially available, NVIDIA is positioning its technology as the foundational layer for a wide range of autonomous driving applications, from robotaxis to advanced driver assistance systems.
Alpamayo 2 Super builds on the architectural strengths of its predecessors while introducing the two-component design of a 32B VLM backbone and a 2.3B diffusion action expert. This design separates reasoning from action generation, allowing each component to be optimized independently. The backbone handles scene understanding and causal reasoning, while the diffusion decoder specializes in generating smooth, feasible trajectories. This division of labor is elegant and practical, and it enables the model to achieve strong performance on both reasoning benchmarks and trajectory accuracy metrics.
The Role of Diffusion Models in Action Decoding
The choice of a diffusion-based action decoder is technically significant. Diffusion models generate outputs by iteratively refining a noisy signal toward a target distribution. For trajectory generation, this means the model can produce diverse, multimodal trajectory distributions rather than a single deterministic path. This is a critical advantage in driving scenarios where multiple valid actions exist — for example, when a vehicle can either merge left or brake to yield. A diffusion decoder can represent both options with associated probabilities, giving downstream planning modules the flexibility to choose the most appropriate action based on additional context.
NVIDIA specifies that the diffusion expert operates over 10 diffusion steps, which is a practical balance between output quality and computational cost. More steps typically produce higher-quality outputs but increase latency. Ten steps represent a reasonable trade-off for a model that produces trajectories at 0.1-second intervals, and the benchmark results suggest this configuration achieves strong spatial accuracy without excessive computational overhead.
Practical Guidance for Engineering Teams Evaluating Alpamayo 2 Super
Engineering teams should approach evaluation with a clear understanding of what this model is and what it is not. Alpamayo 2 Super is a cloud-scale model designed for data center inference. It is not an embedded model for real-time in-car decision-making. The intended use cases are autolabeling, closed-loop simulation, validation, and the generation of training data for smaller models.
Teams should begin with the public notebooks to validate the model against their own data. The six-camera configuration provides a solid reference point, but the architecture supports flexibility in camera count and historical frame depth. Egomotion input is mandatory, so teams need to ensure their data pipeline can provide accurate 3D translation and rotation matrices with timestamps.
For teams working on long-tail scenarios, the CoC trace output is the most valuable feature. It provides a structured explanation of why the model chose a particular action, which is precisely the information needed for scenario-based validation and for building safety cases. The integration with NVIDIA Halos workflows suggests a path toward certification under ISO/PAS 8800, which is increasingly a requirement for deploying autonomous vehicles in regulated markets.
The Distillation Path: From 34B Cloud Model to In-Car Deployment
The gap between cloud-scale training and edge-scale inference is a practical reality that every autonomous driving team faces. Alpamayo 2 Super, at 34B parameters, cannot run on current in-car hardware with acceptable latency and power consumption. Distillation is the standard solution. The process involves training a smaller model to mimic the outputs of a larger teacher model. The smaller model inherits the reasoning capabilities of the teacher through its training objective, but operates with far fewer parameters and lower computational demands.
NVIDIA’s open license explicitly permits this workflow, and the five-output design is well suited to distillation because each output can serve as a training target for a specialized smaller model. A team might distill the VLM backbone into a 3B-parameter model for scene understanding, and separately distill the diffusion decoder into a compact action model. This modular distillation approach preserves the strengths of each component while reducing total computational overhead.
Alpamayo 2 Super is available now, source code is on GitHub, and weights are on Hugging Face under OpenMDW-1.1. The combined package — strong benchmark performance, open licensing, explainable outputs, and a clear integration pathway — positions Alpamayo 2 Super as a practical starting point for teams that are ready to evaluate whether VLA architectures can handle the long-tail events that modular pipelines struggle with. Given the rapid pace of model development in this space, the release also sets a competitive benchmark for what an open-model release should include: permissive licensing, explainable outputs, comprehensive documentation, and a clear roadmap from cloud-scale evaluation to edge-scale deployment.
