NVIDIA Alpamayo 2 Super Open-Source 34B Model Supports Commercial Use, Enhances Long-Tail Scenario Reasoning

NVIDIA released Alpamayo 2 Super on August 4, 2026, a 34B-parameter vision-language-action model built on Cosmos 3 Super Reasoner and post-trained with reinforcement learning. It features open weights under the OpenMDW-1.1 license, designed to provide trajectory planning, causal explanation, and meta-action output for long-tail complex scenarios in robotaxis and autonomous driving.

Alpamayo 2 Super consists of the 32B-parameter NVIDIA Cosmos 3 Super Reasoner and a 2B-parameter diffusion action decoder, post-trained with reinforcement learning. It can output planned trajectories, Chain-of-Causation traces explaining the trajectory, and meta-actions such as yielding or lane changing in a single pass from multi-camera RGB video, text, and ego-vehicle motion history.

Fact Recap

Confirmed facts show that the model achieved a Lingo-Judge score of 79.2 on the LingoQA benchmark, ranking first among nearly 40 models. In NVIDIA tests, it scored 17.0 points higher than Qwen2.5-VL 72B, 15.1 points higher than Gemini 2.5 Pro, and 23.2 points higher than GPT-4o. Closed-loop evaluation scored 1.50 ± 0.13 across 910 scenarios from the PhysicalAI-AV-NuRec dataset on AlpaSim, while open-loop evaluation achieved a 6.4-second minADE of 0.911 meters on 937 challenging samples from the PhysicalAI-AV dataset.

Mechanism Breakdown

Model inputs include 360-degree video from up to seven cameras, language context, and prior motion history. The Reasoner component interprets the scene, while the Action Expert converts internal representations into future ego-vehicle trajectories. Training data includes approximately 115,000 hours of multi-camera driving video, 3.7 million Chain-of-Causation traces, and over 1 billion images. The OpenMDW-1.1 license permits fine-tuning, derivative models, and commercial redistribution, with source code under Apache 2.0.

Industry Impact

For developers, the model provides a single foundation model for trajectory generation, intent prediction, scene understanding, and data annotation, reducing the need to maintain multiple separate models. Enterprise users can obtain weights via Hugging Face and deploy directly in commercial workflows without additional licensing. Upstream and downstream suppliers can leverage the open license to combine proprietary fleet data with the model while retaining control over data and infrastructure.

Comparison and Precedents

Earlier Alpamayo versions were initially research-only, but the OpenMDW license has now been extended to the entire Alpamayo series, making previous versions commercially deployable as well. Alpamayo 2 Super offers the highest reasoning and driving performance within the family, while Alpamayo 1.5 and Alpamayo 1 provide lower-cost options for cloud development and model distillation.

Strategic Assessment

Based on the above facts, the most likely next development is that more autonomous driving teams will adopt the model to generate synthetic training data and teacher outputs for distilling deployment models.

Developers can obtain inference notebooks from GitHub to generate CoC automatic labels and visual question answering outputs directly on multi-camera scenarios. When selecting enterprise solutions, organizations need to assess whether they require a single model that simultaneously handles trajectories, explanations, and meta-actions to match specific deployment costs.