nemotron-3-nano:30b

Nemotron 3 Nano is designed as one model for reasoning and direct-response workloads. It can produce an internal reasoning trace before its final answer, while a chat-template flag can suppress that trace when a concise response is preferred. This makes it practical for applications that need to balance answer quality, transparency, and response presentation across tasks of varying difficulty.

The model uses a hybrid Mixture-of-Experts design with 23 Mamba-2 and MoE layers plus six attention layers. It has 30 billion total parameters, activates 3.5 billion per token, and selects six experts from 128 plus one shared expert in each MoE layer. An NVFP4 variant is also reported to reach up to four times the throughput on Blackwell B200 through Quantization Aware Distillation, making the architecture especially relevant for efficient inference deployments.