Nemotron 3 Nano 30B A3B is a sparse hybrid Mamba-Transformer mixture-of-experts model that pairs 30B total parameters with only about 3B active per token. This design aims to deliver the reasoning capacity of a larger model while behaving in throughput terms much closer to a small dense one, which is useful for teams that want stronger reasoning without paying the latency or cost of a full 30B dense run.

The model is offered with an open-weights license and is positioned for agentic and reasoning workloads, with support for reasoning controls and tool calling alongside very long-context inputs. Practically, it is a good fit for applications that need a balance of quality and efficiency, such as chat assistants, retrieval over large documents, and tool-using agents where the active-parameter footprint matters more than raw parameter count.