Nvidia unveiled a new generation of open models and data libraries, designed for building efficient, transparent and specialised agentic AI systems across industries.

The Nemotron 3 family of models consist of three sizes: Nano, Super and Ultra. Nvidia claims they use a hybrid mixture-of-experts (MoE) design, enabling high throughput and scalability across multi-agent systems.

The models are optimised for low inference costs and high performance, making them ideal for tasks like software debugging, content summarisation, AI assistant workflows and information retrieval.

The company stated the launch of the Nemotron 3 models address the aggressive acceleration of industry adoption of agentic AI applications which require reasoning and tool orchestration.

Nvidia stated organisations are shifting away from single model chatbots to collaborative multi-agent AI systems.

Launching today (15 December), Nvidia explained Nano delivers four times higher throughput than its predecessor and excels in multi-agent tasks with a 1-million-token context window.

The Super model combines 100 billion parameters with 10 billion active parameters. It is designed to provide mid-range intelligence for multi-agent applications.

Finally, Ultra, a large reasoning engine with 500 billion parameters with 50 billion active parameters, delivers more powerful advanced reasoning for complex workflows.

The Super and Ultra models are expected to be available in the first half of 2026

Early adopters include Accenture, Deloitte, EY, Oracle, Palantir, Perplexity, ServiceNow, Siemens, Synopsys, and Zoom. Those companies are using the Nemotron models for AI workflows in manufacturing, cybersecurity, software, media, and communications.

Reuters reported the new family also positions Nvidia to better compete against a growing number of open-source models from companies such as DeepSeek, Alibaba and Moonshot AI.

Meta Platforms has been a big proponent of open source through its Llama large language models (LLMs) but Bloomberg reported it is mulling a shift to towards closed-source models.

.