Published: 6 August 2026
NVIDIA made Alpamayo 2 Super available for commercial use on 5 August 2026, releasing the weights of its 34-billion-parameter vision-language-action model for autonomous driving on Hugging Face under the OpenMDW-1.1 licence. Alpamayo 2 Super scores 79.2 on the LingoQA Lingo-Judge metric, first among nearly 40 models evaluated, and produces a planned trajectory, a causal explanation and a meta-action from a single pass over full-surround camera video. The model was first announced at GTC Taipei on 31 May 2026 and has now shipped.
What is Alpamayo 2 Super and what can it do?
Alpamayo 2 Super is an open vision-language-action model built for level 4 robotaxi development. It pairs a 32-billion-parameter vision-language backbone, built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, with a 2.3-billion-parameter diffusion-based action decoder. Its stated design target is the long tail: rare, multi-agent situations that conventional detection-and-prediction stacks handle poorly.
From one pass over multi-camera video the model emits five outputs: a planned trajectory, a Chain-of-Causation trace explaining that trajectory, a meta-action such as yield, lane change or stop, reasoning auto-labels, and visual question answering with 2D grounding. That combination is what makes it useful operationally, because developers can tie what the model observed to the action it chose. Chain-of-Causation traces integrate with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800.
Inputs are multi-camera RGB video, text, and egomotion history with timestamps. The validated public notebook profiles use six cameras and four historical frames per camera. The trajectory API returns 64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals, each carrying ego-frame XYZ coordinates and a 3×3 rotation matrix.
Alpamayo 2 Super benchmarks and training data
- On LingoQA, Alpamayo 2 Super records a Lingo-Judge score of 79.2 and ranks first among nearly 40 models evaluated.
- In NVIDIA’s testing it beat Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points.
Two further numbers matter for planning work.
- Closed-loop evaluation with AlpaSim across 910 scenarios from the PhysicalAI-AV-NuRec dataset gives an AlpaSim score of 1.50 plus or minus 0.13.
- Open-loop evaluation on 937 challenging samples from the PhysicalAI-AV dataset gives a minADE at 6 seconds of 0.911 metres at the 6.4-second horizon.
Training used roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations, including about 3,700,000 Chain-of-Causation traces, which are structured, causally linked explanations of driving decisions. Image training data exceeds one billion images. The model was tested on a single H100 80GB at 72,115 MiB peak memory, so it is a cloud-scale model intended to be distilled for in-vehicle inference on NVIDIA DRIVE AGX Thor.
What does the OpenMDW-1.1 licence allow?
The weights are released under OpenMDW-1.1, the Linux Foundation’s permissive licence for open model distributions, and the source code is Apache 2.0. The licence covers fine-tuning, derivative models and commercial redistribution. Distilled models can be deployed commercially without further permission from NVIDIA, and model outputs carry no licence conditions.
NVIDIA is applying OpenMDW across the entire Alpamayo family, so earlier releases that were introduced for research and development are now deployable commercially as well. That is the substantive change on 5 August: the model itself was announced at GTC Taipei on 31 May 2026 as expected this summer, and this is the point at which the weights and the commercial terms actually landed.
The Alpamayo family has passed 500,000 downloads on Hugging Face, which NVIDIA cites as making it the most-adopted open reasoning model family for autonomous driving on the platform.
How does Alpamayo 2 Super compare to earlier Alpamayo models?
Alpamayo 2 Super scales to 34 billion parameters against the 10-billion-parameter Alpamayo 1 Nano and Alpamayo 1.5 Nano, a roughly threefold increase that NVIDIA credits with better reasoning, 3D spatial understanding and trajectory prediction in long-tail scenarios.
It also widens the sensor picture. Earlier generations were front-focused; Alpamayo 2 Super uses 360-degree situational awareness across front, side and rear views, which matters for lane changes, merges and intersection crossings. Meta-Action output is new in this generation, adding macro decisions such as yield, lane change and stop on top of trajectories and Chain-of-Causation traces. Reasoning auto-labelling with 2D grounding is also new, and NVIDIA says using the model as an autolabeller on proprietary fleet data compresses annotation cycles from months to days.
Alpamayo 2 Super is designed as a teacher model. A downstream stack built on Alpamayo inherits its reasoning and perception from a single open release, so manufacturers do not each rebuild that infrastructure from scratch.
More information about Alpamayo 2 Super is available now on Hugging Face, and full details on the NVIDIA blog about Alpamayo 2 Super.
