NVIDIA Releases Alpamayo 2 Super: An Open Reasoning Vision Model to Unify Autonomous Vehicle Development

Traditional autonomous driving development relies on multiple independent models to handle trajectory generation, intention prediction, scene understanding, and data labeling. This fragmentation makes it difficult to compare related outputs, investigate model behavior, and reuse representations across workflows. NVIDIA's release of Alpamayo 2 Super directly addresses this pain point by integrating multiple capabilities into a single 34B-parameter open reasoning vision-language-action (VLA) model.

The model consists of a 32B-parameter Cosmos 3 Super Reasoner and a 2B-parameter diffusion-based Action Expert, post-trained with reinforcement learning. The Reasoner interprets multi-camera video, language context, and prior motion history, while the Action Expert converts internal representations into future ego trajectories. This design enables the model to simultaneously output trajectories, reasoning chains, meta-actions, scene question answering, and auto-labels, providing a unified foundation for AV development.

Alpamayo 2 Super's perception covers up to seven cameras for a 360° view and returns multiple complementary outputs: future trajectories, Chain-of-Causation (CoC) reasoning traces, high-level meta-actions, scene question answers, and reasoning auto-labels with 2D grounding. These outputs support four workflows: generating trajectories and CoC reasoning, predicting meta-actions (such as yield, lane change, stop), natural language question answering on multi-camera scenes, and generating CoC auto-labels with 2D grounding.

This multi-task design allows the same foundation model to be reused across different stages of the development pipeline, for example as an offline policy teacher, evaluation critic, data engine, or starting point for new task customization, avoiding the need to maintain separate models for each stage.

In open-loop evaluation, Alpamayo 2 Super achieves a 6.4-second minADE_6 of 0.911 meters on 1,434 challenging samples, a score of 0.433 on the Physical AI AV Reasoning Benchmark, and a LingoQA score of 79.2, ranking first among 37 evaluated models.

On LingoQA, it leads Qwen2.5-VL (72B) by 17.0 points, Qwen3-VL (32B) by 7.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. Additionally, meta-action IoU (lateral 74.59, longitudinal 61.91, lane 73.55), VQA answer similarity 0.652, 2D grounding IoU 0.71, and closed-loop AlpaSim Score 1.50 ± 0.13 all outperform previous models.

The main challenge with open-loop metrics is that they evaluate against fixed pre-recorded futures, failing to capture scene changes after the model's first action. For example, if the ego vehicle changes lanes, the open-loop replay may continue moving neighboring vehicles along recorded trajectories without considering their reactions to the ego vehicle.

Closed-loop simulation executes each predicted action and captures the reactions of surrounding agents when reactive behavior models are included. NVIDIA AlpaSim achieves this by repeatedly rendering observations, providing a more comprehensive evaluation.

Model weights have been released on Hugging Face, and the inference notebook is available on GitHub. The model is released under the OpenMDW-1.1 license, a permissive license from the Linux Foundation for open model distribution, covering fine-tuning, derivative models, and commercial redistribution. Distilled models can be commercially deployed without further NVIDIA licensing, and model outputs carry no license conditions.