skills.rest
Lhhiep-maxcode
Lhhiep-maxcodeCOMMUNITY
Explore all skills by this author →
GitHub

nemotron-super3

Answer questions about NVIDIA Nemotron 3 Super model identity, architecture, training, evaluation, quantization, and release caveats.

Updated Jul 30, 2026
Search Tags:#architecture#evaluation#nemotron#quantization#pretraining#super3
One-click install
npx skills add https://github.com/Lhhiep-maxcode/Nemotron --skill nemotron-super3-lhhiep-maxcode
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: nemotron-super3
Source: https://github.com/Lhhiep-maxcode/Nemotron/tree/main/skills/nemotron-super3
Command: npx skills add https://github.com/Lhhiep-maxcode/Nemotron --skill nemotron-super3-lhhiep-maxcode

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill serves as a reference desk for NVIDIA Nemotron 3 Super, helping users quickly understand the model’s identity, architecture, training pipeline, evaluation results, quantization, and release caveats without digging through multiple files.

Core Features & Use Cases

  • Model Facts and Release Context: Explains the shipped model variant, parameter counts, context length, languages, intended use, and deployment targets.
  • Training and Architecture Guidance: Covers LatentMoE, hybrid Mamba-attention design, pretraining phases, SFT, RL stages, and MTP.
  • Evaluation and Reproducibility: Distinguishes paper results from released recipes, summarizes benchmark behavior, and clarifies what is and is not reproducible from the open repo.
  • Use Case: A researcher can ask what makes Super3 efficient, how long-context training works, or which released checkpoint to use for Hopper versus Blackwell deployment.

Quick Start

Ask for a concise explanation of Nemotron 3 Super, its training stages, and the right source files to read next.

Frequently Asked Questions about nemotron-super3

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

FAQPage Schema
What is the architecture of NVIDIA Nemotron 3 Super and how does LatentMoE work?▼

NVIDIA Nemotron 3 Super uses a hybrid Mamba-attention design with LatentMoE for efficiency. This architecture balances performance with computational overhead during pretraining and inference for specific deployment targets.

How do I reproduce evaluation results from the Nemotron 3 Super open-source recipe?▼

Reproducing Nemotron 3 Super evaluation results requires distinguishing paper claims from open-source recipe behavior. The released recipes summarize benchmark behavior, but specific limitations apply to what is reproducible from the open repository.

Which Nemotron 3 Super checkpoint should I use for Hopper versus Blackwell deployment?▼

Selecting the right Nemotron 3 Super checkpoint for Hopper versus Blackwell deployment depends on the shipped model variant and quantization targets. The model card specifies intended use and deployment targets for these hardware configurations.

What are the pretraining and RL training stages in the Nemotron 3 Super pipeline?▼

The Nemotron 3 Super training pipeline covers multiple pretraining phases, supervised fine-tuning (SFT), reinforcement learning (RL) stages, and multi-token prediction (MTP), building model capabilities progressively.

What are the release caveats and limitations of Nemotron 3 Super quantization?▼

Nemotron 3 Super quantization caveats involve distinguishing paper claims from released recipe behavior. Users must consult the model card to understand context length limits, deployment targets, and what cannot be reproduced from the open repository.

Related Skills

Explore More Education & Research Skills→
Lhhiep-maxcodeLhhiep-maxcode

nemotron-nano3

Answers factual questions about Nemotron 3 Nano architecture, training, evaluation, and deployment from authoritative sources.

Community
Advanced
igamenovoerigamenovoer

krnopt-hw-aware-optimization

Identify GPU compute capability and workload shape for CUDA kernel optimization.

Community
Advanced
tianhao909tianhao909
1

training-llms-megatron

Train large language models with NVIDIA Megatron-Core parallelism strategies.

Community
Advanced
mokbhaimokbhai
1

mcp-builder

Construct MCP servers using official TypeScript and Python SDKs.

Community
Advanced