NVIDIA releases Alpamayo 2 Super model

NVIDIA has introduced Alpamayo 2 Super, a 34-billion-parameter open vision-language-action model designed to streamline autonomous vehicle development. The architecture merges the 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert. Released under the OpenMDW-1.1 license, the model processes 360-degree inputs from up to seven cameras to simultaneously generate future trajectories, Chain-of-Causation reasoning traces, high-level meta-actions, and grounded visual question answering.

The model achieves state-of-the-art results across multiple autonomous driving benchmarks. In open-loop trajectory prediction on 1,434 samples from the Physical AI AV Dataset, it records a 6.4-second minADE_6 of 0.911 meters, outperforming Alpamayo 1.5 Nano's 0.916 meters. On the Physical AI AV Reasoning Benchmark, it scores 0.433, compared to 0.414 for Alpamayo 1.5 Nano and 0.502 for GPT-5.5. It leads the LingoQA benchmark with a score of 79.2, surpassing Qwen3-VL 32B at 72.2, Gemini 2.5 Pro at 64.1, Qwen2.5-VL 72B at 62.2, and GPT-4o at 56.0. In closed-loop testing on 913 reconstructed scenes in AlpaSim, it achieves a score of 1.50 ± 0.13, up from Alpamayo 1.5 Nano's 1.37 ± 0.10.

For scene understanding, Alpamayo 2 Super records 74.59 lateral, 61.91 longitudinal, and 73.55 lane-wise intersection-over-union (IoU) scores on a 94,000-clip meta-action taxonomy. On 8,000 question-answer pairs, it achieves a visual question answering similarity score of 0.652 and a 2D grounding IoU of 0.71, significantly beating Qwen3-VL 32B's scores of 0.450 and 0.17 respectively. For practitioners, this unified design eliminates the need to maintain separate models for labeling, planning, and evaluation. Developers can use the model as an offline policy teacher to distill smaller models for in-vehicle deployment on NVIDIA DRIVE AGX Thor, or run it locally for high-fidelity auto-labeling. Running the model with classifier-free guidance navigation sampling requires two GPUs with at least 70 gigabytes of VRAM.