NVIDIA releases Nemotron 3 Nano 30B-A3B, a hybrid Mamba-MoE reasoning model

NVIDIA published Nemotron 3 Nano 30B-A3B in BF16. The card describes a large language model trained from scratch by NVIDIA as a unified model for reasoning and non-reasoning tasks. It first writes a reasoning trace and then answers, and the trace can be switched off with a flag in the chat template.

The architecture is a hybrid mixture-of-experts: 23 Mamba-2 and MoE layers with 6 attention layers. Each MoE layer has 128 experts plus one shared, with 6 active per token, giving 3.5B active and about 30B total parameters. Languages are English, German, Spanish, French, Italian and Japanese. The card lists many NVIDIA datasets used in training, and describes Nemotron as a family with open weights, data and recipes.

Use is governed by the NVIDIA Nemotron Open Model License, which the card links. In NVIDIA's own table it scores 38.8 on SWE-Bench (OpenHands) against 34.0 for GPT-OSS-20B.

The card gives model dates of September to December 2025; the date used is the Hugging Face repo creation date.

A small active size with its training data listed. The licence is NVIDIA's own.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 4.5: Ungated weights; the card says it is ready for commercial use under NVIDIA's own licence.
• Useful 4: Reasoning on or off from a chat-template flag; scores on agentic coding tasks.
• Notable 4: 831 likes; training datasets listed on the card.
• Novel 3.5: 23 Mamba-2 and MoE layers with 6 attention layers is a distinct hybrid.