Nvidia CEO Jensen Huang has unveiled Alpamayo, an open AI model family aimed at tackling one of the hardest problems in self-driving cars – how to safely handle unpredictable driving scenarios.
“Not only does it take sensor input and activates steering wheel, brakes and acceleration, it also reasons about what action it is about to take,” Huang said at a CES 2026 keynote speech Monday.
For autonomous vehicles, long-tail edge cases such as unusual traffic behavior, unexpected obstacles or unfamiliar road layouts remain a major barrier to scaling beyond limited deployments. Alpamayo addresses this by introducing reasoning-based vision-language-action models that can think through cause and effect step by step, rather than relying solely on pattern recognition.
The models act as large ‘teacher’ systems that help developers train smaller, vehicle-ready models capable of safer decision-making. Combined with open simulation tools and large, diverse driving datasets, Alpamayo enables automakers to test and refine autonomy in complex conditions before real-world deployment, according to the company.
The first car featuring Alpamayo will be a Mercedes-Benz CLA, scheduled to hit U.S. roads by the end of the year.