Alpamayo 2 Super: NVIDIA extends open reasoning to self-driving, amid doubts over scale and hardware requirements

On 4 August, on its corporate blog, NVIDIA announced the availability of Alpamayo 2 Super, presented as the largest model in its family for autonomous vehicles and robotaxis. The system takes images from six cameras, a text prompt and the ego-motion history, and returns a reasoning trace along with a 64-waypoint trajectory covering 6.4 seconds. The announcement describes the new model as "three times the scale" of the earlier 10-billion-parameter versions, yet the official model card published on Hugging Face states a total size of 34 billion — a 32-billion vision-language backbone plus a 2.3-billion diffusion action decoder. The company's two official sources, the blog post and the technical card, do not match. The release comes under the Linux Foundation's OpenMDW-1.1 licence, a permissive choice that allows commercial use and derivative models, setting it apart from the proprietary licences typical of the sector, although the source code remains governed by Apache 2.0.

On performance, NVIDIA reports that Alpamayo 2 Super ranks first on the LingoQA benchmark with a score of 79.2 across nearly 40 models evaluated, ahead of Gemini 2.5 Pro (+15.1 points) and GPT-4o (+23.2 points) under the Lingo-Judge metric. The technical documentation also reports a score of 1.50 ± 0.13 in the AlpaSu closed-loop simulation across 913 scenarios, and a minADE error of 0.911 metres at 6.4 seconds. It is worth being clear about one thing: all of these measurements come from data published by the company itself, which also authored the evaluation tools. As of today there are no independent third-party checks confirming those results outside the environment the vendor controls.

For use on public roads, the sticking point is infrastructure. Technical validation was carried out exclusively on 80 GB NVIDIA H100 GPUs, with peak memory demand above 70 gigabytes — a requirement that, according to MarkTechPost, makes it likely the model would have to be distilled before any in-vehicle integration. No road-test data has been released: current performance is confined to simulation and to the analysis of historical datasets, leaving the system's reliability in real traffic an open question.

Choosing an open licence for a driving system this complex is an interesting signal for the ecosystem, but what counts is the ability to deliver. Promising advanced reasoning on data-centre hardware is not the same as guaranteeing it in a moving car, especially when the official numbers do not line up and the only verification available is the vendor's own. — Olya