NVIDIA ships Alpamayo 2 Super commercial: 34B AV VLA under OpenMDW-1.1

On 4 August 2026 NVIDIA shipped **Alpamayo 2 Super** for commercial use: a 34-billion-parameter vision-language-action model for robotaxi and other autonomous-vehicle work. Weights are on Hugging Face as nvidia/Alpamayo2-Super under **OpenMDW-1.1**. The NVIDIA blog and Hub model card carry the product claims; inference notebooks are on GitHub NVlabs/alpamayo2.

From @nvidia on X (4 August 2026 UTC):

"Meet Alpamayo 2 Super, now commercially available for robotaxis and autonomous vehicles. Built for complex real-world driving, this open reasoning model adds 360° awareness, high-level driving decisions and automated reasoning labels. We built it, so you can build on it."

Architecture on the card is a **32B** VLM backbone on **Cosmos 3 Super** Reasoner plus a **2.3B** diffusion action expert. For a scene it can emit a planned trajectory, a chain-of-causation trace, a meta-action (yield or lane change), auto-labels, and visual Q&A with 2D grounding. Trajectory detail: **64** waypoints from 0.1s through 6.4s at 0.1s steps. Public notebooks use six cameras and four historical frames.

Key facts

• License: weights **OpenMDW-1.1** (fine-tuning, derivatives, commercial redistribution per blog and card); source code **Apache 2.0**. Hub release date **08/04/2026**.
• Family: earlier Alpamayo rows began as R&D; OpenMDW now covers the family for commercial deploy without extra NVIDIA permission. Super is the high-reasoning tier; Alpamayo 1.5 and 1 are leaner cloud or distillation options.
• Vendor evals (NVIDIA testing): LingoQA Lingo-Judge **79.2**; AlpaSim **1.50 ± 0.13** (910 scenarios); open-loop minADE_6 at 6.4s **0.911 m** (1434 samples). Blog Lingo-Judge deltas vs Qwen2.5-VL 72B (+17.0), Gemini 2.5 Pro (+15.1), GPT-4o (+23.2) are NVIDIA-reported.
• Runtime: tested on **1x H100 80GB**; about **72 GiB** peak under the card's seven-camera BF16 profile. Not a laptop claim.
• Card training scale: roughly **115,000** hours multi-camera video and about **3.7M** CoC traces.

How to read this

Commercial OpenMDW weights are the hard fact. Benchmark leads are vendor-reported under NVIDIA protocols, not production SLA or fleet certification. The stated path is cloud-scale reasoner plus distilled in-vehicle models, not the full 34B stack unchanged at the edge. Blog language on Halos and ISO/PAS 8800 is a safety-engineering alignment cue, not an independent robotaxi certificate.

Independence

Frontier Flux is not affiliated with NVIDIA. This piece uses the public blog, Hub card, GitHub presence, and the company X post. Re-check license text and eval tables on the Hub before you plan a commercial stack.