NVIDIA Releases Alpamayo 2 Super For Open Autonomous Vehicle AI Workflows
NVIDIA released Alpamayo 2 Super, an open autonomous vehicle model for trajectory reasoning, labeling, and simulation workflows.
NVIDIA has released Alpamayo 2 Super, an open 34-billion-parameter reasoning vision-language-action model for autonomous vehicle development, alongside practical workflows for testing, labeling, and evaluating driving models. The August 4 technical post frames the release as a way to consolidate tasks that are often split across separate AV systems: trajectory prediction, intent reasoning, scene question answering, grounding, and data labeling.
The model combines a 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert. NVIDIA says Alpamayo 2 Super can process surround-view input from up to seven cameras, use recent vehicle motion history, and produce both a future ego-vehicle trajectory and a Chain-of-Causation trace explaining why the model selected that behavior. That paired output is important for AV teams because it gives reviewers more than a path line; it also gives them a structured explanation they can inspect when a model slows, changes lanes, yields, or fails in a long-tail scene.
Why It Matters
Autonomous driving models are difficult to evaluate when they are tested only against prerecorded futures. In open-loop evaluation, the surrounding traffic does not react to the model's decision, so a single lane change or braking choice may be judged without seeing how the rest of the scene would respond. NVIDIA's post emphasizes closed-loop simulation through AlpaSim, where repeated observations and actions let teams measure failures such as collisions, road departures, and close encounters after the model begins influencing the scenario.
NVIDIA reports a 6.4-second minADE_6 trajectory score of 0.911 meters across 1,434 challenging Physical AI AV Dataset samples, a 0.433 AV reasoning score, 79.2 on LingoQA, and a closed-loop AlpaSim Score of 1.50 plus or minus 0.13 across 913 reconstructed scenes. Those numbers are vendor-reported, but the underlying model weights, inference notebooks, and related code are publicly available through Hugging Face and GitHub, giving researchers a path to inspect and reproduce parts of the workflow.
Developer Access
The release also targets the data side of AV development. Alpamayo 2 Super can generate meta-actions such as stop, yield, accelerate, and change lanes; answer natural-language questions about multi-camera scenes; localize referenced objects with 2D boxes; and propose structured reasoning labels for driving clips. NVIDIA says the model is released under OpenMDW-1.1, a Linux Foundation permissive license for open model distributions, and that distilled models can be deployed commercially without additional permission from NVIDIA.
For the broader AV industry, the notable shift is toward a common open foundation model that can serve as a teacher, evaluator, data engine, and customization starting point. That does not make autonomous driving solved, but it gives teams a more transparent test bed for studying how reasoning models behave before smaller systems are distilled for in-vehicle hardware.
Sources
Cover photo by Abhishek Navlakha on Pexels, used under the Pexels License.
Comments (0)
Leave a Comment