NVIDIA just released a 550-billion-parameter open-source AI model that outperforms every other publicly available American model by a significant margin, while running at speeds that shouldn't be possible for a model this large. The company unveiled Nemotron 3 Ultra during its Computex keynote in Taipei on June 1, 2026, marking a major shift in the open-source AI landscape and raising questions about how NVIDIA achieved such dramatic performance gains.

What Makes NVIDIA's New Model So Dominant?

The performance gap is striking. Nemotron 550B scores 48 on the Artificial Analysis Intelligence Index, a widely used benchmark that measures AI reasoning and knowledge across multiple domains. The next-best American open-source model, Google's Gemma 4, sits at 39. OpenAI's gpt-oss-120b scores 33. Even NVIDIA's own previous flagship, Nemotron 3 Super, only reaches 36. That represents a 9 to 15-point lead over every competitor in the open-source space.

The real shock, however, is speed. This 550-billion-parameter model serves 300 tokens per second on NVIDIA's pre-release endpoint, while comparable Chinese frontier models crawl along at 50 to 100 tokens per second. For context, a token is roughly equivalent to a word or small phrase; 300 tokens per second means the model can generate about 18,000 words per minute. A model this large is not supposed to run this fast, yet NVIDIA has made it happen.

How Does This Fit Into NVIDIA's Broader AI Strategy?

The Nemotron release is part of a larger NVIDIA push into open-source AI infrastructure. On the same day, NVIDIA also launched Cosmos 3, a foundation model designed specifically for physical AI applications like robotics and autonomous vehicles. Cosmos 3 uses a breakthrough mixture-of-transformers architecture, which pairs a reasoning transformer with an expert generation transformer to understand object interactions, motion, and spatial-temporal relationships.

Cosmos 3 is positioned as the world's first fully open omnimodel, meaning it can natively understand and generate text, images, video, ambient sound, and actions with leading physics accuracy. This multimodal capability reduces physical AI training and evaluation cycles from months to days, according to NVIDIA.

What Are the Key Differences Between NVIDIA's New Models?

• Nemotron 550B: A large language model focused on text reasoning and knowledge, designed for general-purpose AI applications and released under a commercial-friendly license.
• Cosmos 3 Super: A physical AI foundation model optimized for post-training robotics and autonomous vehicle models that require the highest physics accuracy and generation quality.
• Cosmos 3 Nano: A lighter-weight version of Cosmos 3 that delivers high-quality video and action reasoning in fractions of a second, suitable for resource-constrained environments.
• Cosmos 3 Edge: Coming soon, this variant is designed for real-time inference at the edge, enabling on-device AI processing without cloud connectivity.

Both model families are available now through Hugging Face, the popular open-source AI model repository, and can be deployed as NVIDIA NIM microservices for production use.

Why Is NVIDIA Betting So Heavily on Open-Source Models?

NVIDIA's aggressive push into open-source AI represents a strategic shift. The company is traditionally known as a hardware manufacturer, but these releases position NVIDIA as a serious player in the software and model development space. By releasing powerful models under commercial-friendly licenses, NVIDIA is building an ecosystem of developers who depend on its infrastructure and tools.

"The big bang of physical AI is just around the corner thanks to breakthroughs in multimodal reasoning language, vision and world models. The Cosmos 3 family of open, frontier omnimodels gives developers a generational leap in ability to build robots, autonomous vehicles and vision AI that perceive, reason, plan and act in the physical world," said Jensen Huang, founder and CEO of NVIDIA.

Jensen Huang, Founder and CEO at NVIDIA

The Cosmos Coalition, a new global collaboration launched alongside Cosmos 3, includes leading AI labs and robotics companies like Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. These partners are working together to advance next-generation world models and contribute research, evaluation techniques, and training resources to the ecosystem.

How to Get Started With NVIDIA's New Open Models

• Access via Hugging Face: Download Nemotron 550B and Cosmos 3 models directly from Hugging Face, the central repository for open-source AI models, where you can review model cards and documentation.
• Try Before Deploying: Experiment with Cosmos 3 on build.nvidia.com to test the model's capabilities for your specific use case before committing to production deployment.
• Customize and Generate Data: Use Hugging Face Diffusers and GitHub resources to customize models for your needs and generate synthetic training data for physical AI applications.
• Deploy as Microservices: Run models as NVIDIA NIM microservices through cloud infrastructure partners including Baseten, CoreWeave, Microsoft Azure, Nebius, Deep Infra, and Classmethod.

Developers building physical AI applications are already adopting Cosmos across industries. Agile Robots, Doosan Robotics, LG Electronics, and Samsung are using it for robotics; Li Auto is leveraging it for autonomous vehicles; and companies like Centific, Fogsphere, and Linker Vision are deploying it for vision AI agents in industrial and smart space applications.

What Does This Mean for the Open-Source AI Landscape?

NVIDIA's dominance in both language models and physical AI foundation models signals a consolidation of power in the open-source AI space. The company is not just providing hardware anymore; it is providing the models, the tools, the deployment infrastructure, and the community frameworks that developers need to build AI systems. This vertical integration gives NVIDIA significant influence over how open-source AI develops.

For developers and enterprises, the availability of these powerful models on Hugging Face means access to state-of-the-art AI capabilities without relying on proprietary cloud APIs. However, the performance gap between NVIDIA's offerings and competitors raises questions about whether other companies can keep pace in the open-source model race.