Overview
- Nvidia introduces a hybrid Mixture‑of‑Experts architecture designed to scale reliable multi‑agent systems while reducing communication overhead and inference costs.
- Nemotron 3 Nano carries roughly 30 billion parameters with up to 3 billion active per token and targets tasks such as debugging, summarization and information retrieval.
- The family also includes Super at about 100 billion parameters with up to 10 billion active per token and Ultra at about 500 billion with up to 50 billion active per token.
- Nvidia says Nano delivers up to four times the token throughput and 60% fewer generated tokens than Nemotron 2 Nano, which the company positions as lowering inference cost.
- Early adopters including EY, Cisco, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys, Zoom, Accenture, Cadence and Cursor are integrating the models into enterprise workflows, with Super and Ultra expected in the first half of 2026 according to the company.