NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1
NVIDIA has launched Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, underneath an open business license. The said design goal is the long-tail occasions: uncommon, multi-agent conditions that typical detection-and-prediction stacks deal with poorly. The mannequin pairs a 32B VLM spine, constructed on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement studying, with a 2.3B diffusion-based motion decoder. From one cross over full-surround digital camera video it emits a deliberate trajectory, a causal rationalization of that trajectory, and a meta-action.
Is it deployable
Yes, and for business use from day one. The weights are launched underneath OpenMDW-1.1, the Linux Foundation’s permissive license for open mannequin distributions; supply code is Apache 2.0. The license covers fine-tuning, spinoff fashions and business redistribution. NVIDIA is making use of OpenMDW throughout your entire Alpamayo household, so earlier releases launched for R&D at the moment are deployable commercially with out extra permission.
Inputs, outputs and coaching knowledge
Inputs are multi-camera RGB video, textual content, and egomotion historical past with timestamps. The validated public pocket book profiles use six cameras and 4 historic frames per digital camera. Egomotion is 3D translation plus a 3×3 rotation matrix, multi-timestep.
The trajectory API returns 64 waypoints spanning 0.1 to six.4 seconds at 0.1-second intervals. Each waypoint carries ego-frame XYZ and a 3×3 rotation matrix.
Training knowledge is roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations. It consists of about 3,700,000 Chain-of-Causation (CoC) traces — structured, causally linked explanations of driving selections. Image coaching knowledge exceeds one billion photographs.
Benchmarks
On LingoQA, Alpamayo 2 Super information a Lingo-Judge rating of 79.2 and ranks first amongst practically 40 fashions evaluated. In NVIDIA’s testing it beat Qwen2.5-VL 72B by 17.0 factors, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2.
Two extra numbers matter for planning work. Closed-loop analysis with AlpaSim on 910 situations from the PhysicalAI-AV-NuRec dataset provides an AlpaSim rating of 1.50 ± 0.13. Open-loop analysis on 937 difficult samples from the PhysicalAI-AV dataset provides minADE₆ at 6.4s of 0.911m.
Five outputs from one mannequin
For every driving scenario, the mannequin produces a trajectory, a CoC hint explaining the choice, a meta-action similar to yield or lane change, reasoning auto-labels, and visible query answering with 2D grounding.
That mixture is what makes the discharge attention-grabbing operationally. Developers can tie what the mannequin noticed to the motion it selected. CoC traces combine with NVIDIA Halos safety-validation workflows and help AI security aligned with ISO/PAS 8800.
Used as an autolabeler on proprietary fleet knowledge, NVIDIA says the mannequin compresses annotation cycles from months to days.