NVIDIA Unveils Expansive Open-Source AI Portfolio to Accelerate Industry Innovation

NVIDIA is significantly broadening the landscape of accessible AI development, today unveiling a comprehensive suite of new open models, extensive datasets, and powerful tools. This release is designed to empower companies across every industry to accelerate the creation and deployment of sophisticated, real-world AI systems.

The newly released models span critical AI domains, including the NVIDIA Nemotron family for agentic AI, the NVIDIA Cosmos platform for physical AI, the new NVIDIA Alpamayo family for autonomous vehicle development, NVIDIA Isaac GR00T for robotics, and NVIDIA Clara for biomedical applications. These contributions are backed by one of the world’s largest collections of open multimodal data, featuring 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures, and 100 terabytes of vehicle sensor data.

Leading technology firms, such as Bosch, CrowdStrike, Palantir, Salesforce, ServiceNow, and Uber, are already integrating and building upon NVIDIA’s open model technologies to enhance their offerings.

Building on the recently introduced Nemotron 3 family, NVIDIA is expanding its Nemotron models to include specialized capabilities for speech, multimodal retrieval-augmented generation (RAG), and enhanced safety features.

• Nemotron Speech: These leaderboard-topping open models, including a new ASR model, deliver real-time, low-latency speech recognition crucial for live captions and advanced speech AI applications. Benchmarks indicate a tenfold performance increase over comparable models.

• Nemotron RAG: Featuring new embed and rerank vision language models (VLMs), this suite provides highly accurate multilingual and multimodal data insights, significantly improving document search and information retrieval processes.

• Nemotron Safety: Designed to bolster the trustworthiness of AI applications, these models now include the Llama Nemotron Content Safety model with expanded language support and Nemotron PII, which accurately detects sensitive personal data.

Bosch is leveraging Nemotron Speech to enable more intuitive driver interactions with vehicles. ServiceNow trains its Apriel model family using open datasets, including Nemotron, to achieve cost-efficient multimodal performance. Additionally, Cadence and IBM are piloting Nemotron RAG models to refine search and reasoning across complex technical documents, while CrowdStrike, Cohesity, and Fortinet are adopting Nemotron Safety models to fortify their AI applications’ trustworthiness. Palantir is integrating Nemotron into its Ontology framework for specialized AI agents, and CodeRabbit uses Nemotron to scale AI code reviews, boosting speed and efficiency.

NVIDIA is also providing developers with open-source datasets, training resources, and blueprints, including the dataset and training code for the Llama Embed Nemotron 8B model. This is complemented by an updated LLM Router for automatically directing AI requests to the optimal model and the dataset used to build the new Nemotron Speech ASR model.

The development of physical AI for robots and autonomous systems demands vast, diverse datasets and models capable of perceiving, reasoning, and acting within complex real-world environments. Robotics is currently the fastest-growing segment on Hugging Face, with NVIDIA’s open robotics models leading platform downloads.

NVIDIA is releasing NVIDIA Cosmos open world foundation models, which integrate humanlike reasoning and world generation capabilities to accelerate the development and validation of physical AI. This includes Cosmos Reason 2, a leading reasoning VLM that enhances robots’ and AI agents’ ability to understand and interact with the physical world, and Cosmos Transfer 2.5 and Cosmos Predict 2.5, which generate large-scale synthetic videos across various environments.

Furthermore, NVIDIA has introduced open models and blueprints specifically for different physical AI embodiments based on Cosmos. Isaac GR00T N1.6 is an open reasoning vision language action (VLA) model tailored for humanoid robots, enabling full body control through Cosmos Reason for superior contextual understanding. The NVIDIA Blueprint for video search and summarization, part of the NVIDIA Metropolis platform, provides a reference workflow for building vision AI agents to analyze vast volumes of video data, enhancing operational efficiency and public safety.

Companies like Salesforce, Milestone, Hitachi, Uber, VAST Data, and Encord are utilizing Cosmos Reason for traffic and workplace productivity AI agents. Franka Robotics, Humanoid, and NEURA Robotics are deploying Isaac GR00T to simulate, train, and validate new robot behaviors before scaling to production.

Developing safe and scalable autonomous driving systems relies on AI that can perceive, reason, and act in intricate real-world scenarios. NVIDIA is introducing NVIDIA Alpamayo, a new family of open models, simulation tools, and extensive datasets designed to advance reasoning-based autonomous vehicle development.

Key components include Alpamayo 1, the first open, large-scale reasoning VLA model for autonomous vehicles (AVs), which enables vehicles not only to understand their surroundings but also to explain their actions. Additionally, AlpaSim is an open-source simulation framework that facilitates closed-loop training and evaluation of reasoning-based AV models across diverse environments and edge cases.

NVIDIA is also releasing Physical AI Open Datasets, comprising over 1,700 hours of driving data collected across a wide range of geographies and conditions. This data specifically covers rare and complex real-world edge cases vital for advancing reasoning architectures in autonomous vehicles.

To reduce costs and accelerate treatment delivery, NVIDIA is launching new Clara AI models that bridge the critical gap between digital discovery and real-world medicine. These models are designed to help researchers develop treatments that are safer, more effective, and easier to produce.

• La-Proteina: Enables the design of large, atom-level-precise proteins for research and drug candidate development, offering new tools to study previously intractable diseases.

• ReaSyn v2: Ensures that AI-designed drugs are practical to synthesize by integrating a manufacturing blueprint directly into the discovery process.

• KERMT: Provides high-accuracy computational safety testing early in development by predicting how a potential drug will interact with the human body.

• RNAPro: Unlocks the potential of personalized medicine by predicting the complex 3D shapes of RNA molecules.

Complementing these models, an NVIDIA dataset of 455,000 synthetic protein structures is available to help AI researchers build more accurate AI models for drug discovery and biological research.

NVIDIA’s open models, data, and frameworks are now accessible on platforms such as GitHub and Hugging Face, as well as from various cloud, inference, and AI infrastructure providers, and directly via build.nvidia.com. This flexible access ensures developers can readily leverage these resources. Many of these models are also available as NVIDIA NIM microservices, facilitating secure and scalable deployment on any NVIDIA-accelerated infrastructure, from edge devices to cloud environments. Further details can be found by watching NVIDIA Live at CES.