Late at night on August 4, Jensen Huang, CEO of NVIDIA, announced the official launch of the open-source model Alpamayo 2 Super for autonomous vehicles, supporting self-driving taxis, trucks, shuttles, delivery vehicles, tractors and other vehicle types in long-tail scenarios. Alpamayo 2 Super was first released at the GTC 2026 Taipei conference on June 1 this year. Huang commented that "Alpamayo marks the transition of automobiles from pure driving to safety reasoning". Alpamayo 2 Super is a 32-billion-parameter Vision-Language-Action (VLA) large model built based on Cosmos 3 Super Reasoner. Unlike traditional end-to-end "black box" models, it explains every decision in natural language and adds "Meta-Action" output, directly outputting high-level driving strategies such as yielding, emergency pullover, roundabout detour, and temporary parking, rather than only underlying control quantities like steering wheel angle and braking force. It also features 2D reasoning-based automatic annotation, compressing manual annotation workload from several months to several days, and a "teacher-student" distillation architecture that uses the 32-billion-parameter model as a teacher to distill a smaller student model for the NVIDIA DRIVE AGX Thor platform. Additionally, it upgraded from front camera perception to 360° environment perception, covering front, side and rear directions.