NVIDIA Open-Sources Autonomous Driving Inference Model at NeurIPS 2025

At NeurIPS 2025, NVIDIA unveiled a comprehensive suite of new open models, datasets, and tools spanning autonomous driving, speech AI, and safety research, further solidifying its leadership in open digital and physical AI development.

The company also received recognition from Artificial Analysis’s newly launched Open Index, which ranks NVIDIA’s Nemotron series among the most transparent model ecosystems.

NVIDIA introduced DRIVE Alpamayo-R1, which it describes as “the world’s first open inference Vision-Language-Action (VLA) model for autonomous driving.”

Bryan Catanzaro, Vice President of Applied Deep Learning Research at NVIDIA, explained that the model integrates chain-of-thought reasoning with path planning to support research into complex road scenarios and Level 4 autonomy.

According to NVIDIA, AR1 progressively decomposes driving scenes, evaluates potential trajectories, and leverages contextual data to determine optimal routes. Part of its training data is available through NVIDIA’s Physical AI Open Datasets, and the model itself can be accessed on GitHub and Hugging Face.

Built on NVIDIA Cosmos Reason, AR1 is customizable for non-commercial research. NVIDIA noted that reinforcement learning significantly enhances the model’s reasoning capabilities during post-training, outperforming its pre-trained version. The company also released AlpaSim, an open framework for evaluating AR1.

Furthermore, NVIDIA expanded the Cosmos ecosystem in its “Cosmos Handbook” with new tools and workflows, offering step-by-step guidance for model post-training, synthetic data generation, and evaluation.

New systems based on Cosmos include LidarGen, a world model for generating LiDAR data; Omniverse NuRec Fixer, which corrects artifacts in neural reconstructions; Cosmos Policies, which transforms video models into robotic policies; and ProtoMotions3, a framework for training physically simulated digital humans and robots.

Industry partners—including Voxel51, 1X, Figure AI, Foretellix, Gatik, Oxa, PlusAI, and X-Humanoid—are already leveraging Cosmos World Foundation Models. Researchers from ETH Zurich presented NeurIPS work demonstrating how Cosmos models generate coherent 3D scenes.

In the realm of digital AI, NVIDIA launched new models and datasets under the Nemotron and NeMo families. These include MultiTalker Parakeet, a speech recognition model designed for multi-speaker environments; Sortformer, a calendar-aware model; and Nemotron Content Safety Reasoning, which applies domain-specific safety rules.

NVIDIA also open-sourced the Nemotron Content Safety Audio Dataset for detecting unsafe audio content. Additionally, it released synthetic data and reinforcement learning tools, including NeMo Gym—a reinforcement learning environment—and the NeMo Data Designer library, now available under the Apache 2.0 license.

Partners such as CrowdStrike, Palantir, and ServiceNow are utilizing Nemotron and NeMo tools to build specialized agentic AI solutions.