Nvidia has used its annual GTC conference to release a broad collection of open artificial intelligence models, meaning software that developers can freely access, modify and build on, covering some of the most consequential areas where AI is being applied.
The clearest way to understand what is on offer is to think about the three distinct problems Nvidia is trying to solve: making AI that can hold conversations and take actions, making AI that can control physical machines, and making AI that can accelerate medical research.
On the conversational side, Nvidia has expanded its Nemotron family of models, which are designed to power AI agents, systems that do not just answer questions but reason through problems and carry out multi-step tasks.
The headline addition, Nemotron 3 Ultra, is designed to deliver performance comparable to the most advanced AI models while running five times more efficiently, which matters because lower computing costs make powerful AI cheaper to deploy at scale.
Two companion models handle voice: one combines audio, vision and language understanding, and another merges speech recognition with text generation for real-time voice conversations of the kind increasingly used in customer service and virtual assistants.
Companies including CrowdStrike, the cybersecurity firm, and Perplexity, the AI search service, are already using these models.
For physical AI, the goal is giving robots and self-driving vehicles a better understanding of the real world, with Nvidia's Cosmos 3 and Isaac GR00T models designed to help machines perceive their environment, reason about what they see and decide how to act.
GR00T, in particular, is a foundation model for humanoid robots, providing a general-purpose base that robotics companies can train for specific tasks rather than building from scratch.
Nvidia has also previewed the next generation of GR00T, due by the end of the year, which suggests the pace of development in physical AI is accelerating rapidly.
In healthcare, Nvidia's BioNeMo platform is being used to design protein binders, molecules engineered to attach to specific proteins in the body, which is an early step in developing new drugs.
Pharmaceutical companies, including Novo Nordisk have already been testing designs generated by these models.
Related reading
- NVIDIA launches AI factory reference design and digital twin blueprint
- Nvidia and Dassault Systèmes combine simulation software and AI physics tools to speed up industrial design
- NVIDIA Nemotron 3 Super boosts agentic AI throughput 5x
Nvidia also collaborated with Google DeepMind and others to add 1.7 million high-confidence protein structure predictions to the AlphaFold database, expanding one of science's most important open resources for understanding how biology works at a molecular level.
All of these models are available through platforms including GitHub and Hugging Face, with Nvidia offering packaged deployment options for organisations that want to use them in commercial products without building infrastructure from the ground up.
The recap
- NVIDIA unveils Nemotron, Cosmos, Isaac and BioNeMo model families
- Added about 30 million protein complex predictions to AlphaFold
- Select models available on GitHub, Hugging Face and build.nvidia.com