NVIDIA Alpamayo 2 Super: Open 34B Reasoning VLA Model for Autonomous Vehicle Development

What happened

Source factNVIDIA announced Alpamayo 2 Super, an open 34-billion-parameter reasoning vision-language-action (VLA) model for autonomous vehicle development. The model combines the 32-billion-parameter NVIDIA Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert and is post-trained with reinforcement learning. It accepts multi-camera video from up to seven cameras covering 360 degrees, along with language context and prior motion history, and outputs future trajectories, Chain-of-Causation reasoning traces, meta-actions, grounded question answering, and reasoning auto-labels.

Source factThe model is released under the Linux Foundation permissive license OpenMDW-1.1, which covers fine-tuning, derivative models, and commercial redistribution; distilled models can be deployed commercially without further permission from NVIDIA, and model outputs carry no license conditions. Weights are available on Hugging Face and inference notebooks on GitHub.

Evidence assessment

Source factQuantitative results reported by NVIDIA include a 6.4-second minADE_6 of 0.911 m on 1,434 challenging samples, an AV reasoning score of 0.433, a LingoQA score of 79.2 ranking first among 37 models, and an AlpaSim closed-loop score of 1.50 ± 0.13 on 913 reconstructed scenes. Meta-action accuracy on 94K clips yields 74.59 lateral IoU, 61.91 longitudinal IoU, and 73.55 lane-wise IoU. VQA on 8K pairs achieves 0.652 answer similarity and 0.71 grounding IoU, outperforming Qwen3-VL 32B at 0.450 and 0.17 respectively.

What happened

Source factThe blog demonstrates four workflows: generating trajectories with CoC traces and evaluating them open-loop and closed-loop; predicting meta-actions; performing visual question answering with 2D grounding; and generating CoC auto-labels from user clips. The model can auto-label clips without future trajectory by predicting its own future motion.

What is actually new

AI analysisThe actual novelty lies not in any single capability but in their integration into one open-weight model: a single 34B VLA that jointly reasons about scene context, produces an action, explains it, and can label data. The 360-degree seven-camera setting and grounded VQA with bounding boxes are also notable. Prior models separate trajectory generation, intent prediction, and labeling; Alpamayo 2 Super collapses these stages into a common representation.

Constraint shift

AI analysisThis release shifts a practical constraint in AV development: the need to maintain and align separate models for planning, reasoning, and labeling. A single open foundation model can serve as policy teacher, evaluation critic, data engine, and customization starting point. The permissive OpenMDW-1.1 license and unencumbered model outputs lower legal and cost barriers for commercial adoption, while auto-labeling promises to compress annotation cycles from months to days.

Why it matters

AI analysisBecause autonomous driving requires interpreting rare and complex scenarios, an open model that jointly predicts trajectories and reasons in natural language can improve interpretability, data curation, and safety validation. The LingoQA result and grounding IoU suggest the model's scene understanding is strong enough to support semi-automated labeling and closed-loop evaluation. This may accelerate the shift from modular AV stacks to foundation-model-based pipelines.

Opportunity signal

AI hypothesisIf the reported numbers hold up under independent evaluation, Alpamayo 2 Super could become a default open baseline for AV perception-planning research and for commercial data annotation services. The open license and end-to-end workflows create a business opportunity around fine-tuned specialized models and evaluation services, rather than around proprietary model access.

Evidence assessment

AI analysisAll quantitative evidence comes from a single NVIDIA blog post with internal evaluation sets; no external validation, statistical significance tests, or disaggregated error analysis are provided. The open-loop and closed-loop metrics are indicative but not comparable across prior methods without standardized evaluation protocols. The model's compute requirements (e.g., 70 GB VRAM for classifier-free guidance) may limit reproducibility.

What would change my mind

AI hypothesisI would revise the assessment if independent reproductions show materially different benchmark scores, if the open license contains undisclosed restrictions, if closed-loop performance degrades on reactive benchmarks with more diverse agents, or if the model's auto-labels require heavy human correction in practice. Verification on public datasets like nuScenes or Waymo Open would strengthen confidence.