# NVIDIA Opens Its 32-Billion-Parameter Driving Model for Commercial Use

Alpamayo 2 Super triples the parameter count of its predecessor and, in NVIDIA's own testing, leads the LingoQA driving-reasoning benchmark against nearly 40 models.

- Published: 2026-08-05T05:32:18.568Z
- Canonical: https://polylog.news/ai/2026-08-05/nvidia-opens-its-32-billion-parameter-driving-model-for-comm
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [NVIDIA Blog](https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/), [NVIDIA Newsroom](https://nvidianews.nvidia.com/news/nvidia-alpamayo-2-super-robotaxis), [The Next Web](https://thenextweb.com/news/nvidia-alpamayo-2-open-reasoning-driving-model)

NVIDIA has moved Alpamayo 2 Super, [its largest open driving model](https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/), from preview to commercial availability. The model is a vision-language-action system with 32 billion parameters, up from 10 billion in the prior Alpamayo release, and it adds 360-degree perception and what NVIDIA calls Meta-Actions. Inference code ships through GitHub and the weights through Hugging Face.

The design target is rare road situations. Ordinary highway and intersection handling is largely solved by conventional perception and motion-prediction systems. What remains are unusual configurations that were never densely represented in training data. NVIDIA's argument is that a model which reasons explicitly about a scene, in text, generalizes to those cases better than one that maps sensor input directly to trajectories.

On benchmarks, NVIDIA reports that Alpamayo 2 Super [ranks first on LingoQA](https://nvidianews.nvidia.com/news/nvidia-alpamayo-2-super-robotaxis) among close to 40 evaluated models, and that under the Lingo-Judge metric it outscores Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points. Those comparisons come from NVIDIA's own runs, and the baselines are general-purpose multimodal models rather than driving-specialized systems, which makes the margin look larger than a like-for-like test would. LingoQA measures the quality of driving-scene question answering, not closed-loop driving safety.

The commercial licensing is the substantive change. NVIDIA says the Alpamayo family has passed 500,000 downloads on Hugging Face, and until now that adoption could not carry into shipped vehicle software.

## What this means

An open, commercially usable 32-billion-parameter driving model narrows the distance between companies with proprietary autonomy stacks built over a decade and newer entrants, particularly Chinese and European robotaxi and truck programs that can fine-tune downloadable weights on their own fleet data. The channel is distribution rather than raw capability. NVIDIA gains because its own inference hardware becomes the default target for anyone who adopts the model. The parties exposed are autonomy startups whose differentiation was the base perception and reasoning model rather than fleet data or operations.

## What to watch

- Whether an independent group reproduces the LingoQA margins against driving-specialized baselines rather than general multimodal models.
- Whether any named robotaxi operator commits to Alpamayo in a production stack, which is the test of whether open weights change deployment and not just research.
- Closed-loop safety results, since scene question answering does not measure intervention rates on road.