NVIDIA Alpamayo 2 Super Generates Trajectories and Reasoning Traces

NVIDIA released Alpamayo 2 Super, a 34-billion-parameter vision-language-action model that generates trajectories, reasoning traces, and structured auto-labels. The model combines a 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert, and processes up to seven cameras for 360-degree perception. Outputs include Chain-of-Causation (CoC) reasoning traces, meta-actions, VQA with 2D grounding, and auto-labels.

Performance and Comparisons

On open-loop evaluation using 1,434 Physical AI AV Dataset samples, Alpamayo 2 Super achieves a minADE_6 of 0.911 meters, improving on Alpamayo 1.5 Nano's 0.916 meters. The reasoning score on the Physical AI AV Reasoning Benchmark is 0.433, higher than Alpamayo 1.5 Nano's 0.414 but lower than GPT-5.5's 0.502. The LingoQA score is 79.2, which NVIDIA states is 17.0 points above Qwen2.5-VL 72B. In closed-loop simulation on 913 reconstructed scenes, the AlpaSim Score is 1.50 ± 0.13, up from Alpamayo 1.5 Nano's 1.37 ± 0.10. Meta-action prediction reaches 74.59% lateral IoU, 61.91% longitudinal IoU, and 73.55% lane-wise IoU on 94K internal clips. VQA achieves 0.652 answer similarity and 0.71 grounding IoU on 8K question-answer pairs. CoC auto-labeling similarity to expert annotations is 0.652 on 8K internal clips.

Deployment and Inference

Model weights are hosted on Hugging Face under the ID nvidia/Alpamayo2-Super, with inference notebooks available on GitHub. The model is released under the OpenMDW-1.1 license, a Linux Foundation permissive license. Distilled versions may be deployed commercially without additional NVIDIA permission, and model outputs carry no license restrictions.

For inference, NVIDIA suggests using torch.autocast with bfloat16, top-p 0.98, temperature 0.6, 10 diffusion steps, and one trajectory sample. The batch size should start at 1 due to memory intensity. For closed-loop evaluation with AlpaSim, the command uses uv run alpasim_wizard with deploy=local, topology=2gpu, and driver=alpamayo2. The AlpaSim driver configuration requires two GPUs with at least 70 GB VRAM for classifier-free guidance navigation. AlpaSim uses four cameras at 10 Hz, with a context length of 4 frames per camera. Route configuration disables waypoint commands, making navigation language-only. The trajectory optimizer is disabled by default for Alpamayo 2 Super.

Developers are encouraged to join the Alpamayo developer forum for support.