At #NeurIPS2025 we’re delivering new open-source tools for building, training, and deploying smarter AI systems, medical research, AV development, edge inference, and much more over 70 papers, talks, and workshops.
Here’s a few #NVIDIAResearch highlights Build safer AV systems with Alpamayo R1: Our open VLA with chain-of-thought reasoning shows 35% fewer close encounters +12% better planning accuracy. Open weights + AlpaSim coming to Github/HF. Scale your AV + robotics training data with the Cosmos “cookbook,” showing you how to augment datasets with weather, lighting, or terrain variations using open World Foundation Models.
Create high-quality synthetic datasets using our Nemotron Open-Source model. Generate privacy-safe datasets for AI-safety reasoning, use NeMo Gym Training environments for reinforcement learning, and our new multi-speaker speech models.
More Tools for better models with fast edge inference to deeper reasoning: Audio Flamingo 3 for audio, Minitron-SSM pruning for compression, Jet-Nemotron hybrid NAS and Nemotron-Flash SLMs for speed, and ProRL for reasoning and capability gains.Learn how to use the open models, datasets, and research we’re releasing
https://blogs.nvidia.com/blog/neurips-open-source-digital-physical-ai/