Furthering its deep commitment to open source, NVIDIA is unveiling new open-source AI technologies for language, robotics and biology — contributing to an open ecosystem that broadens access to AI and fuels U.S. innovation.

These open technologies will empower developers worldwide and strengthen economic growth through efficient reasoning, high-fidelity world generation and interactive physical AI systems accelerated on NVIDIA infrastructure.

The new open models, data and tools are part of the NVIDIA Nemotron family for AI reasoning, the NVIDIA Cosmos platform for physical AI, NVIDIA Isaac GR00T for robotics and NVIDIA Clara for biomedical AI.

NVIDIA is contributing these models, data and training frameworks to Hugging Face to make AI research and development more accessible. As a top contributor to Hugging Face, with more than 650 open models and 250 open datasets now available, NVIDIA continues to expand access to cutting-edge AI resources for the global developer community.

“Open models are catalysts to AI innovation, making AI accessible, transparent and responsible,” said Clément Delangue, CEO of Hugging Face. “NVIDIA’s contributions to the open model ecosystem, commitment to open research for AI and Hugging Face’s ecosystem will empower millions of developers to build advanced AI — together and in the open.”

NVIDIA Nemotron Brings Ultra-Efficient Reasoning to Specialized AI Agents

AI agents use multiple models to see, retrieve, generate and reason. The latest open models in the NVIDIA Nemotron family unify these capabilities, enabling developers to build specialized, intelligent agents.

• Nemotron Nano 3 uses a hybrid mixture-of-experts architecture to improve reasoning throughput for areas like software development, customer service and IT support.
• Nemotron Nano 2 VL provides document intelligence, image reasoning and video analysis.
• Nemotron Parse extracts text and tables from documents to take actionable insights.
• Nemotron Safety Guard adds culturally aware multilingual moderation capabilities, detecting harmful content across 23 safety categories in nine languages.
• Nemotron retrieval-augmented generation (RAG) models now feature advanced document extraction and unified retrieval across text, images, audio and video data sources.

NVIDIA also released new open-source Nemotron datasets — including on multimodal training, multilingual personas and privacy-preserving synthetic personal information — for specialized model development. In addition, new NVIDIA NeMo tools, including NeMo Data Designer for synthetic data generation and NeMo-RL for advanced post-training and reinforcement learning, give developers greater control over model customization.

Leading Software Companies Building Agentic AI on Nemotron

Open-source models enable developers to build and adapt AI for their domains, transforming software into agentic systems that use tools, not just provide them. Leading software and services companies are building agentic software platforms using NVIDIA Nemotron.

Building on the success of the Apriel Nemotron model family that’s post-trained with NVIDIA and ServiceNow-provided data, ServiceNow introduced its new Apriel 2.0 multimodal reasoning model that brings intelligence to cross-enterprise workflows for every industry, including regulated sectors like financial services, healthcare and telecom.

“Open models are driving the next wave of enterprise transformation,” said Joe Davis, executive vice president of platform engineering and AI at ServiceNow. “Apriel 2.0 represents one of the first open-weight multimodal reasoning models built for the enterprise — combining text, document and data understanding to power real-world workflows across industries. Together with NVIDIA, we’re pairing reasoning transparency with performance efficiency, bringing explainable, secure and scalable AI to sectors where trust and compliance matter most.”

Additional leading software companies are adopting NVIDIA’s latest reasoning models to power their next-generation AI applications:

• Palantir is making Nemotron available via its Ontology in Foundry and AIP to support an integrated technology stack with NVIDIA for operational AI.
• Cadence is accelerating chip designer productivity with its JedAI Platform and Nemotron open technologies.
• CrowdStrike is building autonomous, continuously learning AI agents into its Agentic Security Platform with NVIDIA Nemotron.
• PayPal is boosting the throughput and cost efficiency of its AI services by 50% with Nemotron.
• Synopsys is collaborating with NVIDIA to develop chip-design agents with the NVIDIA NeMo Agent Toolkit and Nemotron open models and data, used by NVIDIA engineers.
• Zoom is utilizing Nemotron for customized agentic capabilities to align with the specific workflows of its customers.

Cosmos and Isaac GR00T Open Models and Data for Physical AI and Robotics

To accelerate the training of robotic systems with humanlike reasoning and cognition, NVIDIA introduced major updates to its open models for physical AI, including Cosmos world foundation models and Isaac GR00T robot foundation models:

• Cosmos Predict 2.5: Unifies three models into one for rapid world simulation, generating 30-second videos from a single frame.
• Cosmos Transfer 2.5: Produces higher-quality, photorealistic data from 3D scenes at one-third the size of Cosmos Transfer 1.
• Cosmos Reason: A reasoning vision language model, now available as an NVIDIA NIM microservice, for advanced multimodal understanding.
• Cosmos Dataset Search: Enables instant retrieval of training scenarios, cutting post-training cycles from months to days.
• Isaac GR00T N1.6: Enhances reasoning, generalization and whole-body control for humanoid robots.

NVIDIA also released the world’s largest open-source dataset for physical AI, now featuring 1,700 hours of multimodal driving sensor data from across the U.S. and Europe as well as GR00T training data, which has risen to the top 10 most-downloaded Hugging Face datasets of all time.

Leading companies — including Agility Robotics, Amazon Robotics, Figure AI, Skild AI, Milestone Systems and Uber — are adopting Cosmos or Isaac GR00T N models to generate synthetic data, teach robots new behaviors and deploy real-world and physical AI agents at scale.

NVIDIA Clara Open Models for Healthcare and Life Sciences

New open models joining NVIDIA Clara, a family of models, tools and recipes built for accelerating scientific discovery, analyzing medical images and more, include:

• Clara CodonFM: Learns the rules of RNA to reveal how changes in its code can improve the design of therapies and medicine. NVIDIA will contribute open models like CodonFM to the Chan Zuckerberg Initiative’s virtual cells platform, accelerating open-source collaboration and model evaluation.
• Clara La-Proteina: Creates 3D protein structures atom by atom at twice the length and complexity of previous models, enabling the design of better medicines, enzymes and materials.
• Clara Reason: A vision language model enabling chain-of-thought reasoning for radiology and medical imaging to advance explainable AI medical research.

Get Started With NVIDIA Open Models

Select NVIDIA Nemotron and Cosmos open models, trained on NVIDIA DGX Cloud, are available on build.nvidia.com, Hugging Face, OpenRouter and Microsoft Azure AI Foundry, and are coming soon to Google Vertex AI Platform and other cloud service providers. NVIDIA Clara and Isaac GR00T are available on Hugging Face.

The models are available as NVIDIA NIM microservices for secure, scalable deployment on DGX Cloud or any NVIDIA-accelerated infrastructure for maximum privacy and control.

Learn more about how NVIDIA and partners are advancing AI innovation in the U.S. by watching the NVIDIA GTC Washington, D.C., keynote by Huang.