nemotron-nano3

Retrieve repo-cited facts about Nemotron 3 Nano architecture and training.

1.8k|373|Updated Oct 3, 2025
One-click install
npx skills add https://github.com/NVIDIA-NeMo/Nemotron --skill nemotron-nano3
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: nemotron-nano3
Source: https://github.com/NVIDIA-NeMo/Nemotron/tree/main/skills/nemotron-nano3
Command: npx skills add https://github.com/NVIDIA-NeMo/Nemotron --skill nemotron-nano3

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Nemotron 3 Nano users need fast, authoritative answers about the model’s architecture, training data, recipes, evaluation results, quantization, and deployment behavior without wading through a large repository.

Core Features & Use Cases

  • Paper-first retrieval: resolves questions using the Nano3 tech report chunks for architecture, pretraining, SFT, RL (RLVR/GRPO/RLHF), evaluation, and safety/alignment construction.
  • Public-recipe grounding: maps what the public repo’s Nano3 stage recipes expose, including where they match the paper “shape” versus where they do not reproduce proprietary mixtures.
  • Model-card checkpoint guidance: answers which released checkpoints exist (Base/BF16/FP8), what they’re for, and how to interpret context and reasoning controls for deployment.
  • Handoff boundary: when the user’s goal becomes procedural (build, fine-tune, reproduce pipelines, customize to hardware/data), the skill directs them to /nemotron-customize.

Quick Start

Use the nemotron-nano3 skill when you ask, in plain English, for facts about how Nemotron 3 Nano works (for example: “How many experts are activated per token, and what is the active-parameter count?”).

Frequently Asked Questions about nemotron-nano3

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
What is Nemotron 3 Nano and what architecture does it use?

Nemotron 3 Nano is a language model with a specific architecture and active-parameter count per token. Its design details, including activated experts, are grounded in cited chunks from the official Nano3 technical report.

How was Nemotron 3 Nano trained using SFT and reinforcement learning?

Nemotron 3 Nano training involved supervised fine-tuning (SFT) and reinforcement learning methods like RLVR, GRPO, and RLHF. Authoritative facts about these pretraining and alignment recipes are retrieved directly from the model card and paper.

Which released Nemotron 3 Nano checkpoints support FP8 quantization?

Released Nemotron 3 Nano checkpoints include Base, BF16, and FP8 variants. The FP8 checkpoint supports specific quantization deployment behaviors, with context and reasoning controls detailed in the official model card.

What do evaluation benchmarks say about Nemotron 3 Nano throughput?

Evaluation benchmarks and throughput claims for Nemotron 3 Nano are documented in the tech report. The skill retrieves these specific performance metrics with citation-by-file-path grounding from the paper chunks.

Can I use the public Nemotron 3 Nano repo recipes to reproduce the exact proprietary training mixtures?

The public Nemotron 3 Nano repo recipes expose the paper's general shape but do not reproduce proprietary training mixtures. The skill explicitly separates paper claims from these public recipe reproduction limits.

How do I build and fine-tune a Nemotron 3 Nano pipeline on custom hardware?

To build, fine-tune, or reproduce Nemotron 3 Nano pipelines on custom hardware, you are directed to the nemotron-customize skill. This skill handles factual questions, routing procedural build procedures to that dedicated tool.