Touting its status as the “world’s largest contributor to open-source AI,” Nvidia Corp. is doubling down on open artificial intelligence development with a broad expansion of its open model ecosystem and the launch of a global collaboration to build next-generation frontier AI systems.

The company today introduced new models across several vertical market domains and unveiled a set of partnerships with AI labs and developers to collaborate on building open foundation models.

The initiatives, announced at Nvidia’s Global Technology Conference in San Jose, reflect the company’s growing effort to position open models as a core layer of the AI economy. Open models only make up about 20% of all model usage, but cost an average of 87% less in production, according to MIT researchers.

Beyond chat

At the center of the announcement is an expansion of Nvidia’s Nemotron family of models, which support the emerging generation of AI agents that can perform complex tasks rather than simply answer questions.

The new Nemotron 3 lineup includes multimodal models that integrate language, vision and voice capabilities. Nvidia said the models are designed to support natural conversations, complex reasoning and the ability to extract insights from multiple forms of data.

Among the new offerings is Nemotron 3 Ultra, which Nvidia said delivers frontier-level performance while achieving five times greater throughput efficiency using the NVFP4 floating-point format developed by Nvidia for use on its Blackwell platform. The model is intended for AI-native applications such as coding assistants, enterprise search and automated workflows.

“It’s been pretrained completely in NVFP4, is the most intelligent and has two times the max throughput of the current best open base model,” said Kari Briski, senior vice president of generative AI software at Nvidia.

Other models include Nemotron 3 Omni, which combines audio, vision and language understanding so AI systems can interpret documents and video simultaneously, and Nemotron 3 VoiceChat for real-time conversational interaction combining speech recognition, language model processing and text-to-speech capabilities into a single system.

To improve reliability, Nvidia also introduced Nemotron safety models and an agentic retrieval pipeline designed to detect unsafe content and improve response accuracy in multimodal environments.

Nvidia said a growing ecosystem of enterprise software companies is already adopting the models. Automation Anywhere Inc., CodeRabbit Inc., CrowdStrike Inc., AnySphere Inc., Distyl AI Inc., The San Francisco AI Factory Inc., Genspark Inc., Perplexity AI Inc. and ServiceNow Inc., are deploying Nemotron models to power AI-driven applications. LangChain Inc. has also integrated the models into its agent development framework.

Nvidia is also expanding its open model strategy into robotics and autonomous systems with three new offerings. Cosmos 3 is a foundation model designed to simulate complex environments and support physical reasoning for robots and autonomous machines.

Isaac GR00T N1.7 is a vision-language-action model for humanoid robots. Alpamayo 1.5, improves reasoning and navigation capabilities in autonomous vehicles. Nvidia said organizations including HCL Technologies Ltd., Johnson & Johnson MedTech, Milestone Systems A/S and the Toyota Research Institute are already using Cosmos while LG Electronics Inc., Neur Robotics GmbH and Noble Machines Inc. are adopting GR00T to scale humanoid robotics deployments.

Scientific discovery

In healthcare and life sciences, the BioNeMo platform, provides AI tools for modeling biological systems. A new generative model called Proteina-Complexa is designed to accelerate drug discovery by generating protein binders that can attach to target molecules.

Nvidia also collaborated with Google DeepMind, the  European Molecular Biology Laboratory’s European Bioinformatics Institute and Seoul National University to expand the AlphaFold protein structure database with tens of millions of predicted protein complex structures.

The additional data is expected to help researchers identify new drug targets and better understand disease biology, the company said.

In a related move, Nvidia announced the Nemotron Coalition, a partnership aimed at advancing open frontier models through shared expertise, datasets and computing infrastructure.

“Building frontier models demands significant time, expertise and compute, which is a major investment most organizations can’t make alone,” Briski said. “That’s why we’re launching the Nemotron Coalition.”

Initial members include Black Forest Labs Inc, AnySphere, LangChain, Mistral AI SAS, Perplexity, Reflection AI Inc., Axonwise Private Ltd. and Thinking Machines Lab.

The coalition’s first project will be a new open base model codeveloped by Nvidia and Mistral AI and trained on Nvidia’s DGX Cloud infrastructure. It’s expected to be released to the open ecosystem and will serve as the foundation for the upcoming Nemotron 4 family of models.

Image: Nvidia