NVIDIA has introduced a new slate of open models, data, and developer tooling aimed at accelerating AI adoption across industries ranging from enterprise software and cybersecurity to robotics, autonomous driving, and biomedicine. NVIDIA said the releases expand its open model universe with updates spanning the Nemotron family for agentic AI, the Cosmos platform for physical AI, the new Alpamayo family for autonomous vehicle development, the Isaac GR00T robotics model line, and new Clara models for healthcare and life sciences.
Alongside the models, NVIDIA said it is contributing open-source training frameworks and a large collection of open multimodal datasets intended to speed training and evaluation of real-world AI systems. The company highlighted the scale of its open resources, including 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures and 100 terabytes of vehicle sensor data, positioning the package as a broad foundation for work in language, robotics, scientific research and autonomous vehicles.
On the agentic AI side, NVIDIA said it is releasing new Nemotron models focused on speech, retrieval-augmented generation and safety. Nemotron Speech includes an automatic speech recognition model designed for low-latency transcription and captioning, while Nemotron RAG adds embed and rerank vision-language models aimed at improving multilingual and multimodal document search. For safety, NVIDIA said Nemotron Safety models now include a Llama Nemotron Content Safety model with expanded language support and a Nemotron PII model designed to detect sensitive data.
NVIDIA also emphasized physical AI and robotics development, citing growing interest in robotics on model hubs and pointing to Cosmos open world foundation models designed to help systems perceive, reason and act in complex environments. It said Cosmos Reason 2 is a reasoning vision-language model meant to improve physical-world understanding, while Cosmos Transfer 2.5 and Cosmos Predict 2.5 are focused on generating synthetic video across environments and conditions. NVIDIA also highlighted Isaac GR00T N1.6, a reasoning vision-language-action model designed for humanoid robots, and noted a reference workflow for video search and summarization as part of its Metropolis platform.
For autonomous vehicles, NVIDIA said its new Alpamayo family combines open models, simulation tools and datasets built for reasoning-based autonomy. The company said Alpamayo 1 is an open reasoning vision-language-action model intended to help vehicles interpret surroundings and explain actions, while AlpaSim is an open-source simulation framework for closed-loop training and evaluation. NVIDIA also pointed to “Physical AI Open Datasets” that include more than 1,700 hours of driving data across geographies and conditions to capture rare and complex edge cases.
In healthcare and life sciences, NVIDIA said it is launching new Clara AI models to close the gap between digital discovery and real-world medicine. The company highlighted La-Proteina for atom-level protein design, ReaSyn v2 to incorporate manufacturability into AI-driven drug discovery, KERMT to predict drug interactions for early-stage safety testing, and RNAPro to predict RNA 3D structures, alongside a dataset of synthetic protein structures.
NVIDIA said the open models and supporting resources are available through major developer distribution channels including GitHub, Hugging Face and its own developer platforms, and that many are also offered as NVIDIA NIM microservices for deployment on NVIDIA-accelerated infrastructure from edge to cloud.
