NVIDIA has announced a comprehensive release of open-source AI models, opening the door to significant advancements in AI development across multiple domains such as language processing, robotics, and synthetic world generation. Central to this release are the Nemotron models aimed at creating specialized AI agents available for various smart applications.
Nemotron Models for AI Innovation
The newly launched Nemotron series includes:
<ul>
<li>Nemotron Nano 3: A 32B MoE (Mixture of Experts) model featuring 3.6B active parameters tailored for efficient reasoning tasks.</li>
<li>Nemotron Nano 2 VL: A 12B multimodal model designed for comprehensive document and video understanding.</li>
<li>Nemotron Parse 1.1: A document parser with 1B parameters.</li>
<li>Nemotron RAG Suite: A suite of models optimized for retrieval-augmented generation pipelines.</li>
<li>Llama 3.1 Nemotron Safety Guard: A model built for multilingual content moderation.</li>
</ul>
These models integrate cutting-edge architectures such as hybrid Mamba-Transformer and support innovations in quantization techniques like FP4/FP8 for efficient model deployment. They are open-source and accessible on platforms like Hugging Face, along with datasets and training recipes.
Beyond Language to Physical and Biological AI
Additionally, NVIDIA has diversified its AI offerings with other model families:
<ul>
<li>Cosmos: Focused on physical AI and world simulation, these models provide tools for generating synthetic data crucial for training robots and autonomous systems when integrated with platforms like Isaac Sim and Omniverse.</li>
<li>Isaac GR00T: Catering to humanoid robotics needs.</li>
<li>Clara: Dedicated to advancements in biomedical applications.</li>
</ul>
The overarching release constitutes over 650 models and 250 datasets, already being leveraged by industry leaders such as ServiceNow, Palantir, and PayPal. These models showcase capabilities in multimodal reasoning, rapid video generation, and protein structure prediction, available as NIM microservices for seamless deployment on NVIDIA infrastructure or cloud solutions.
Synthetic Worlds and Simulation
The Cosmos models (Predict 2.5 and Transfer 2.5) further enhance the capability to create synthetic training environments. These updates provide highly detailed photorealistic data, crucial for physical AI training, significantly cutting costs and time associated with gathering real-world data. Companies such as Skild AI, Serve Robotics, and Zipline are advancing robotic model training and validation in simulated environments before real-world application.
NVIDIA’s commitment to open-source accessibility fosters an ecosystem of AI innovation, paving the way for wide-ranging applications from efficient multilingual moderation to pioneering physical AI systems.

Collections

NVIDIA released over 650 open-source AI models and 250 datasets spanning language processing, robotics, and synthetic world generation. The Nemotron series includes models for reasoning, multimodal understanding, document parsing, and RAG pipelines, featuring hybrid Mamba-Transformer architectures and FP4/FP8 quantization. Additional model families include Cosmos for physical AI simulation, Isaac GR00T for humanoid robotics, and Clara for biomedical applications. These models are available on Hugging Face and as NIM microservices, already being used by ServiceNow, Palantir, and PayPal for applications ranging from content moderation to robot training in synthetic environments.

Introducing NVIDIA's Open-Source AI Models: Enhancing Language, Robotics, and Specialized AI Agents