Nemotron 3 Nano 30B A3B

Nemotron 3 Nano 30B A3B is a large language model developed and trained from scratch by NVIDIA as a single, unified system that can handle both reasoning and non-reasoning tasks. Rather than relying on separate models for different jobs, it first produces a reasoning trace and then delivers a final answer, and users can toggle that reasoning behavior through a flag in the chat template. Disabling the trace gives a leaner response style, while leaving it on typically produces stronger results on harder prompts that genuinely benefit from step-by-step thinking, making the model adaptable across everyday chat and more analytical workloads.

The architecture is what gives the model its efficiency profile. It uses a hybrid Mixture-of-Experts design that mixes 23 Mamba-2 and MoE layers with 6 attention layers, where each MoE layer contains 128 routed experts plus 1 shared expert and activates 6 experts per token. That structure yields roughly 3.5B active parameters out of 30B total, so it behaves like a much smaller model at inference time while retaining the capacity of a larger one. The generous context window makes it well suited to long-document summarization, code repositories, multi-turn analytical sessions, and any task where maintaining coherence across very large inputs matters.