NVIDIA Releases Alpamayo 2 Super Open Model for Autonomous Driving

NVIDIA released Alpamayo 2 Super, a 34 billion parameter vision-language-action (VLA) model, on an open commercial license, targeting complex driving scenarios. This model is engineered to address "long-tail events," which are infrequent, multi-agent situations that traditional detection and prediction systems struggle to manage effectively. Alpamayo 2 Super integrates a 32 billion parameter VLM backbone, built upon NVIDIA's Cosmos 3 Super Reasoner and enhanced with reinforcement learning. This backbone is coupled with a 2.3 billion parameter diffusion-based action decoder. The model processes full-surround camera video in a single pass to generate a planned trajectory, a causal explanation for that trajectory, and a meta-action. The model is immediately deployable for commercial use under the OpenMDW-1.1 license, a permissive license from the Linux Foundation for open model distributions, with its source code available under the Apache 2.0 license. This license permits fine-tuning, the creation of derivative models, and commercial redistribution. NVIDIA is extending the OpenMDW license across its entire Alpamayo family, meaning earlier research and development releases are now commercially usable without requiring additional permissions. The input data for Alpamayo 2 Super includes multi-camera RGB video, text, and historical egomotion data with timestamps. Publicly available notebooks demonstrate configurations using six cameras and four historical frames per camera. Egomotion data comprises 3D translation and a multi-timestep 3x3 rotation matrix. The trajectory API outputs 64 waypoints, spaced at 0.1-second intervals over a 6.4-second period, with each waypoint specifying ego-frame XYZ coordinates and a 3x3 rotation matrix. The training dataset for Alpamayo 2 Super consists of approximately 115,000 hours of multi-camera driving video, annotated with egomotion and trajectory information. This dataset also includes around 3,700,000 Chain-of-Causation (CoC) traces, which are structured, causally linked explanations of driving decisions. The image training data exceeds one billion images. In benchmarks, Alpamayo 2 Super achieved a Lingo-Judge score of 79.2 on the LingoQA benchmark, ranking first among nearly 40 evaluated models. NVIDIA's internal testing indicated that Alpamayo 2 Super outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. The release signifies a step towards more robust and explainable AI for autonomous systems, particularly in handling unpredictable real-world driving conditions.