Briefs · Taipei time (GMT+8)

NVIDIA introduces Alpamayo 2 Edge, a reasoning model for cars, and Halos SAFE, a framework for judging self-driving safety

一句重點:The goal is for Alpamayo 2 Edge to read the road and map out a route in real time, running 10 times per second on DRIVE AGX Thor, the in-car platform. It is a small model: it has 2 billion parameters and builds on Cosmos (NVIDIA press release).

發生什麼事

Halos SAFE is about proving that a car is safe enough. It produces safety evidence that can be traced back to its source, by checking behavior against scenarios and metrics fixed ahead of time. Those checks don't stop at simulation: they carry on into real-vehicle testing and keep running as the fleet is monitored after it reaches public roads. TÜV Rheinland, IVEX and dSPACE helped draw up how the assessment works (NVIDIA press release).

Other new tools include AlpaJudge, which grades driving reasoning, and Cosmos 3 Nano variants for producing more synthetic data. Furthest out is Alpamayo Agentic Factory, due in early 2027, whose point is to save integration work: teams have had to wire up simulation-based evaluation, post-training and data preparation on their own, and this open blueprint hands the job of linking those steps to agents (NVIDIA press release).

01Sources