Nemotron 3 Nano 30B A3B

Nemotron 3 Nano 30B A3B is a large language model trained from scratch by NVIDIA and presented as a unified system capable of handling both reasoning and non-reasoning tasks. When prompted, the model first produces an internal reasoning trace and then delivers a final answer, with the reasoning behavior togglable through a flag in the chat template. Disabling reasoning produces faster but slightly less accurate responses on harder problems, while leaving it on generally yields higher-quality solutions.

Under the hood, the model pairs a hybrid Mixture-of-Experts backbone with selective state-space components, alternating 23 Mamba-2 and MoE layers with 6 attention layers. Each MoE layer routes from 128 experts plus a shared expert, activating 6 per token, which keeps the active footprint at 3.5B parameters against a total of 30B. This design aims to deliver reasoning-grade quality while keeping inference efficient, making the model well suited for developers who want controllable thinking behavior, open-weight deployment, and a compact active parameter budget in a single artifact.