NVIDIA used NeurIPS to broaden its open-source ecosystem for digital and physical Artificial Intelligence, publishing new models, datasets and developer tools and presenting over 70 papers, talks and workshops. The company noted recognition from the Artificial Analysis Open Index, which rates the NVIDIA Nemotron family of open technologies among the most open in the Artificial Intelligence ecosystem based on license permissibility, data transparency and technical detail availability.
A central release is NVIDIA DRIVE Alpamayo-R1 (AR1), described as the world’s first industry-scale open reasoning vision language action model for autonomous vehicle research. AR1 integrates chain-of-thought Artificial Intelligence reasoning with path planning to address nuanced driving scenarios and to support level 4 autonomy. The model reasons through possible trajectories, uses contextual data to choose routes and exposes reasoning traces that can inform future planning decisions. AR1 is built on NVIDIA Cosmos Reason and can be post-trained with reinforcement learning, which the company reports significantly improved reasoning compared with the pretrained model. AR1, related training data subsets and the AlpaSim evaluation framework will be available on GitHub and Hugging Face, and some training and evaluation data are published in the NVIDIA Physical AI Open Datasets.
NVIDIA also expanded its Cosmos developer resources with the Cosmos Cookbook and multiple Cosmos-based examples, including LidarGen for lidar generation, Omniverse NuRec Fixer for neurally reconstructed data correction, Cosmos Policy for robot policy extraction and ProtoMotions3 for training simulated humanoids. On the digital side, Nemotron additions include MultiTalker Parakeet and Sortformer for multi‑speaker speech tasks, Nemotron Content Safety Reasoning and a Nemotron Content Safety Audio Dataset for cross-modal safety, plus NeMo Gym and the NeMo Data Designer Library, now open-sourced under Apache 2.0, to support reinforcement learning and synthetic dataset generation. NVIDIA named partners using these tools and invited attendees to the Nemotron Summit, taking place today from 4-8 p.m. PT, while research highlights at the conference include Audio Flamingo 3, which handles audio segments up to 10 minutes and achieves state-of-the-art results on over 20 benchmarks, Minitron-SSM pruning Nemotron-H 8B to 4 billion parameters with 2x faster inference throughput, Jet-Nemotron, Nemotron-Flash and ProRL. Events run through Dec. 7 in San Diego.
