NVIDIA Nemotron 3: How Open AI Models Are Powering the Next Wave of Agentic, Physical and Healthcare AI

Posted on 22nd Jul 2026 06:03:38 in Artificial Intelligence, Machine Learning

Tagged as: NVIDIA, Nemotron 3, open source AI, AI agents, physical AI, healthcare AI, robotics

NVIDIA has made a significant stride in democratizing artificial intelligence with its expanding family of open models. With the launch of the Nemotron 3 family of open models, data, and libraries, the company has introduced a breakthrough hybrid latent mixture-of-experts architecture that helps developers build and deploy reliable multi-agent systems at scale. These open models are designed to power the next wave of agentic AI, physical AI systems, and healthcare breakthroughs, giving organizations the tools to build specialized intelligence that fits their unique needs.

Nemotron 3: Open Models Built for Agentic AI at Scale

The centerpiece of NVIDIA's open model strategy is the Nemotron 3 family, which introduced a breakthrough hybrid latent mixture-of-experts (MoE) architecture. Available in three sizes — Nano (30 billion parameters, 3 billion active), Super (100 billion parameters, 10 billion active), and Ultra (500 billion parameters, 50 billion active) — the Nemotron 3 models deliver an unprecedented combination of efficiency and accuracy for building multi-agent AI systems.

Nemotron 3 Nano achieves up to 4x higher token throughput than its predecessor and reduces reasoning-token generation by up to 60%, significantly lowering inference costs. With a 1-million-token context window, it can handle long, multistep tasks with greater accuracy. Independent benchmarking by Artificial Analysis ranked it as the most open and efficient model among its size class with leading accuracy.

Beyond the core reasoning models, NVIDIA expanded the Nemotron family with omni-understanding multimodal models. Nemotron 3 Omni integrates audio, vision, and language understanding, enabling AI agents to extract insights from videos and documents with high efficiency. Nemotron 3 VoiceChat supports real-time conversations where AI listens and responds simultaneously, combining automatic speech recognition, large language model processing, and text-to-speech in a single system. Safety models and a retrieval pipeline strengthen trustworthy multimodal systems by detecting unsafe content across text and images.

Leading enterprises are already deploying Nemotron models in production. ServiceNow introduced Apriel 2.0, an open-weight multimodal reasoning model built on Nemotron for cross-enterprise workflows. CrowdStrike is building autonomous AI agents into its Agentic Security Platform. Palantir is making Nemotron available through its Foundry and AIP platforms. CodeRabbit, Cursor, Perplexity, Factory, and Distyl are among the many companies integrating Nemotron to power advanced agentic applications across industries ranging from cybersecurity to software development.

Physical AI Breakthroughs: Cosmos, Isaac GR00T, and Alpamayo

NVIDIA's commitment to open models extends deep into the physical world with significant updates to its Cosmos world foundation models, Isaac GR00T robot foundation models, and Alpamayo autonomous vehicle models. These technologies are designed to help robots, autonomous vehicles, and other physical AI systems perceive, reason, and act in real-world environments.

NVIDIA Cosmos 3 represents a milestone as the first world foundation model to unify synthetic world generation, physical AI reasoning, and action simulation. Cosmos Predict 2.5 condenses 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 its predecessor. Cosmos Reason, a reasoning vision language model, is now available as an NVIDIA NIM microservice for advanced multimodal understanding.

The Isaac GR00T N1.7 model is a reasoning vision-language-action (VLA) model purpose-built for humanoids, now commercially viable for real-world deployment. NVIDIA also previewed GR00T N2, based on DreamZero research, which helps robots succeed at new tasks in new environments more than twice as often as leading VLA models. GR00T N2 currently ranks No. 1 on MolmoSpaces and RoboArena for generalist robot policies.

For autonomous vehicles, NVIDIA Alpamayo 1.5 — a reasoning VLA model — supercharges autonomous driving with navigation guidance, prompt conditioning, flexible multi-camera support, and configurable camera parameters. Companies including Agility Robotics, Amazon Robotics, Figure AI, Skild AI, Milestone Systems, Uber, LG Electronics, and NEURA are adopting Cosmos or Isaac GR00T models to generate synthetic training data, teach robots new behaviors, and deploy physical AI at scale.

NVIDIA also released the world's largest open-source dataset for physical AI, featuring 1,700 hours of multimodal driving sensor data from across the United States and Europe. The GR00T training data has risen to the top 10 most-downloaded Hugging Face datasets of all time, reflecting the massive demand for open resources in robotics and autonomous systems research.

Healthcare AI: BioNeMo and Clara Models Accelerating Scientific Discovery

NVIDIA's open model initiative is transforming healthcare and life sciences through the BioNeMo platform and Clara model family. These open technologies enable researchers to model, design, and simulate biological systems at scale, accelerating everything from drug discovery to medical imaging analysis.

The Proteina-Complexa model, part of the BioNeMo platform, accelerates protein drug discovery. NVIDIA developed this model alongside Google DeepMind, EMBL's European Bioinformatics Institute, and Seoul National University, and released a new open dataset comprising millions of AI-predicted protein complex predictions. This collaboration exemplifies how open models can pool expertise across industry and academia to tackle fundamental challenges in biomedical research.

New models joining the NVIDIA Clara family include Clara CodonFM, which learns the rules of RNA to reveal how changes in its genetic code can improve the design of therapies and medicines. NVIDIA is contributing open models like CodonFM to the Chan Zuckerberg Initiative's virtual cells platform, accelerating open-source collaboration and model evaluation in the life sciences.

Healthcare organizations are putting these models to work. Novo Nordisk, Viva Biotech, and Manifold Bio are leveraging NVIDIA's healthcare AI models for drug discovery and biomedical research. Abridge is customizing Nemotron to build the first foundation model purpose-built for clinical conversations. Heidi Health is delivering frontier-quality outcomes in clinical documentation without requiring frontier-scale compute power.

The Strategic Value of Open Models for Enterprise AI

Behind NVIDIA's open model push is a clear philosophy: open models give enterprises something closed models cannot — full control to customize, inspect, and improve AI against their own business needs. Public benchmarks measure general capability, but business-specific evaluation lets teams test against their own data, workflows, and definitions of accuracy.

The cost advantages are substantial. Arcee AI achieved inference costs of roughly 90 cents per million output tokens by post-training Nemotron on the NVIDIA Blackwell platform — approximately 20x cheaper than comparable closed frontier models. LangChain tuned its Deep Agents harness for Nemotron 3 Ultra and achieved top agent accuracy among open models at approximately 10x lower cost per run than leading closed alternatives.

NVIDIA also launched the Nemotron Coalition, bringing model builders and developers together to improve Nemotron through shared data, evaluations, and domain expertise. Organizations from AI Singapore to YTL AI Labs are using Nemotron to build sovereign models that serve billions of people in their native languages while aligning with local cultures and values.

With more than 650 open models and 250 open datasets on Hugging Face, NVIDIA has established itself as one of the leading contributors to the open-source AI ecosystem. As Kari Briski, NVIDIA's vice president of generative AI software, stated: "Open source AI has become a global force for innovation. From biology and scientific discovery to robotics and autonomous machines, NVIDIA open model families extend intelligence beyond language, enabling developers worldwide to build intelligent agents and power breakthroughs across digital and physical industries."

Sources

• NVIDIA Newsroom — NVIDIA Expands Open Model Families to Power the Next Wave of Agentic, Physical and Healthcare AI
• NVIDIA Newsroom — NVIDIA Debuts Nemotron 3 Family of Open Models
• NVIDIA Blog — NVIDIA Launches Open Models and Data to Accelerate AI Innovation
• NVIDIA Blog — Nemotron Labs: How Open Models Give Enterprises and Nations AI They Can Trust, Control and Customize