Nvidia releases Alpamayo 2 edge weights (local AV reasoning) (developer.nvidia.com)

Nvidia has launched Alpamayo 2 Edge, a cutting-edge adaptive reasoning vision-language-action (VLA) model designed specifically for real-time operations in autonomous vehicles (AVs) on the NVIDIA DRIVE AGX Thor platform. This new model addresses the crucial challenge of fitting complex reasoning capabilities into constrained onboard systems by optimizing performance and reducing model size. Featuring a 2B-parameter Nemotron-based reasoning backbone and techniques like offline distillation, closed-loop reinforcement learning, and adaptive reasoning, Alpamayo 2 Edge generates Chain-of-Causation (CoC) reasoning that enhances driving trajectories while conserving computational resources during routine driving scenarios. The introduction of Alpamayo 2 Edge is significant for the AI/ML community as it establishes a strong foundation for onboard reasoning in autonomous driving and showcases Nvidia’s commitment to enhancing the capabilities of AV systems. With improvements in closed-loop driving performance by 48.9% over its predecessor, stakeholder feedback highlights remarkable advances in lane-keeping ability and faster trajectory planning. Furthermore, the integration of AlpaJudge—a human-calibrated evaluation tool—allows developers to assess reasoning quality and optimize training workflows, ensuring models retain high performance and accuracy even as their size is reduced for onboard deployment. This advancement marks a substantial stride forward in making intelligent, adaptive AV systems more efficient and effective.