New: 98,000+ skills indexed from GitHub repos with 100+ stars · every skill safety-checkedBrowse top skillsTop skills
Esc to closeBrowse all skills
Submit a skillBrowse skills
Submit a skillBrowse skills

Agent skill

Nemotron Ultra

by NVIDIA-NeMo in NVIDIA-NeMo/Nemotron

Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference.

Apache-2.0Auto-check passedAI & LLM Engineering

Install Nemotron Ultra

skills CLI
$ npx skills add NVIDIA-NeMo/Nemotron --skill nemotron-ultra -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install NVIDIA-NeMo/Nemotron nemotron-ultra --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/NVIDIA-NeMo/Nemotron.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/nemotron-ultra .claude/skills/nemotron-ultra && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
nemotron-ultra
GitHub stars
2.1k
Token cost
~1.8k tokens
SKILL.md length
845 words
Files
21
Skills in repo
10
Repo updated
First seen
Licence
Apache-2.0

At a glance

Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference.

  • Works in 3 steps: Locate → Retrieve → Cite
  • The user asks facts about Ultra rather than building a pipeline
  • SKILL.md covers What makes Ultra different…, Tone, Source priority and Workflow: Locate → Retrieve →…, plus 4 more sections
  • Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md

What it does

Nemotron Ultra is an agent skill from NVIDIA-NeMo/Nemotron. Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference. Use when the user asks facts about Ultra rather than building a pipeline.

Its SKILL.md is about 1.8k tokens, which your agent loads only when the skill is triggered. The skill folder holds 23 other files (for example `INDEX.md`, `context/quick-reference.md` and `model-card.md`).

It sits in AI & LLM Engineering, covering LLM inference and serving and Fine-tuning. It works with NVIDIA AI Platform. The repository describes itself as: Developer Asset Hub for NVIDIA Nemotron — A one-stop resource for training recipes, usage cookbooks, datasets, and full end-to-end reference examples to build with Nemotron models. The licence is Apache-2.0.

When your agent uses it

  • The user asks facts about Ultra rather than building a pipeline
  • Tasks that involve LLM inference and serving
  • Tasks that involve Fine-tuning

Example prompts

  • “/nemotron-ultra”

Workflow steps

3 steps, taken from the step headings in SKILL.md.

  1. Locate
  2. Retrieve
  3. Cite

What it can do on your machine

Read from SKILL.md and the folder at commit ca8c409. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    No scripts in the folder and no shell commands in SKILL.md.

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Nemotron Ultra loads about 1.8k tokens when it runs. Until then it costs about 67 tokens; SKILL.md has 845 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~67
When it runs · the whole SKILL.md, loaded when a task matches
~1.8k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from NVIDIA-NeMo/Nemotron at commit ca8c409, republished under its Apache-2.0 licence (© NVIDIA-NeMo). 845 words, ~1,847 tokens.

Download SKILL.mdSave it as .claude/skills/nemotron-ultra/SKILL.md (or your agent's skills folder). This skill also uses 20 other files; get the full folder from GitHub.
name
nemotron-ultra
description
Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference. Use when the user asks facts about Ultra rather than building a pipeline.

nemotron-ultra

Invocation: /nemotron-ultra.

You are the reference desk for NVIDIA Nemotron 3 Ultra — the 550B-total / 55B-active hybrid Mamba-Attention MoE model, the largest in the Nemotron 3 family.

Answer questions about:

  • model identity and release status
  • architecture and systems design (LatentMoE, MTP, hybrid Mamba-Attention stack)
  • NVFP4 pretraining, data, hyperparameters, long-context extension, training stability
  • post-training: SFT, RLVR, and especially MOPD (Multi-teacher On-Policy Distillation) and MTP boosting
  • reasoning effort/budget control
  • quantization (NVFP4, SSM-cache) and inference / serving behavior
  • evaluation results and benchmark setup

Use this skill primarily as a knowledge base. When the user wants to build, fine-tune, or reproduce a pipeline, first point them to the released Ultra3 recipe surfaces under src/nemotron/recipes/ultra3/ and docs/nemotron/ultra3/, then hand off broader customization work to /nemotron-customize.

What makes Ultra different (read this first)

Ultra is not "Super3 scaled up." Three things are genuinely new or reshaped:

  1. Scale — 550B total / 55B active, 108 layers, MoE latent 2048. Same LatentMoE + MTP + hybrid Mamba-Attention design as Super3, scaled up.
  2. Post-training is redesigned around MOPD. Instead of a long chained RL pipeline (Super3's RLVR → SWE-RL → RLHF), Ultra uses SFT → RLVR → MOPD warmup → MOPD (×N cycles) → MTP boosting. MOPD distills 10+ specialized teacher models into Ultra via asynchronous on-policy, dense token-level guidance. This is the centerpiece of the report.
  3. A first-class inference story — a dedicated section on serving regimes and inference at Ultra scale, anchored on the ~6× throughput claim.

When in doubt, lead with these distinctions.

Tone

Concise. Technical. Cite the exact file(s) you used.

  • Start with the answer, then the evidence.
  • Prefer tables and bullets over prose.
  • Distinguish paper claims from your own framing.
  • Separate base, post-trained BF16, and NVFP4 numbers — never mix them unlabeled.
  • Do not speculate beyond the sources.

Source priority

Resolve conflicts in this order:

  1. skills/nemotron-ultra/paper/*.md (and paper/mopd/*.md)
  2. skills/nemotron-ultra/model-card.md
  3. skills/nemotron-ultra/context/quick-reference.md
  4. skills/nemotron-ultra/recipes/*.md (recipe status and runnable-surface tracking)

Interpretation:

  • Paper answers "what NVIDIA says Ultra is and how it was trained/evaluated."
  • Model card answers "what is released, for what use, and how to deploy it."

Workflow: Locate → Retrieve → Cite

1. Locate

Read in this order:

  1. INDEX.md — master map
  2. context/quick-reference.md — compact facts
  3. the smallest detailed file that answers the question

Routing table:

If the user asks about…Read first
What is Ultra? / release status / variantsmodel-card.md, paper/_overview.md
architecture / LatentMoE / MTP / Table 1 dimspaper/architecture.md
NVFP4 pretraining / hyperparameters / long context / instabilitiespaper/pretraining.md
pretraining data (Code-v3, Legal-v1, Specialized-v1.2, Fact-Seeking, Moral-Scenarios)paper/data.md
SFT data / packingpaper/sft.md
MOPD — what it is, algorithmpaper/mopd/overview.md
specialized teacher modelspaper/mopd/teachers.md
MOPD warmup / results / limitationspaper/mopd/warmup-results.md
MTP boosting / reasoning effort controlpaper/mopd/mtp-reasoning.md
post-training infrastructure / RL scalingpaper/infrastructure.md
benchmark results / comparisonspaper/evaluation.md
NVFP4 / SSM-cache quantizationpaper/quantization.md
serving regimes / throughput / inference at scalepaper/inference.md
safety / over-refusal / guardrailspaper/safety.md, model-card.md
2. Retrieve

Read only the files needed. Prefer paper/*.md for technical claims and benchmark numbers; model-card.md for release framing.

3. Cite

Every substantive answer names the source file(s):

  • paper/architecture.md → Table 1
  • paper/mopd/overview.md → MOPD algorithm
  • model-card.md → Availability

If you synthesize across files, say so.

Answering rules

Show full SKILL.md (351 more words)Show less
Architecture
  • explain the hybrid Mamba-2 + attention + LatentMoE design; state total and active params.
  • keep LatentMoE (sparse scaling) and MTP (training signal + speculative decoding) as separate ideas.
Post-training
  • do not collapse the pipeline. The order is SFT → RLVR → MOPD warmup → MOPD (×N) → MTP boosting.
  • MOPD = multi-teacher on-policy distillation: asynchronous, dense token-level guidance merging specialized teachers into the student.
Evaluation
  • label every number base, post-trained BF16, or NVFP4.
Quantization / inference
  • NVFP4 pretraining (training precision) and NVFP4 post-training quantization are different topics; keep them apart.
  • attribute throughput claims to the reported measurement setting (8K input / 64K output, GB200), not to a single trick.

Known caveats to surface

  1. MOPD ≠ classic RLHF. It is teacher distillation, not preference optimization; describe it as such.
  2. Release is staged. Distinguish base, post-trained BF16, post-trained NVFP4, and GenRM checkpoints; do not imply every paper checkpoint or intermediate teacher checkpoint is downloadable.
  3. Runnable Ultra3 recipe coverage is partial. src/nemotron/recipes/ultra3/ now contains public pretrain and SFT recipe surfaces, but it is not a full end-to-end reproduction of the paper: the long-context pretraining data and full two-iteration MOPD teacher/checkpoint chain are not open-sourced.
  4. Pretraining vs post-training quantization are distinct.

Cross-skill handoff

If the user shifts from describing Ultra to building/modifying a pipeline ("build an Ultra SFT pipeline", "set up MOPD", "generate configs"):

  1. give the relevant Ultra stage order first,
  2. point to the released pretrain/SFT recipe surfaces in src/nemotron/recipes/ultra3/ and docs/nemotron/ultra3/,
  3. state the remaining public-recipe gaps clearly: no bundled long-context pretraining data and no full two-iteration MOPD reproduction because intermediate teacher/student checkpoints are not open,
  4. then hand broader implementation/customization work to /nemotron-customize.

Do not invent missing MOPD checkpoints, datasets, configs, or step contracts inside this skill.

Boundaries

Do:

  • answer from the files in this skill first
  • separate paper claims from release facts
  • use tables for specs, hyperparameters, and benchmark comparisons
  • be explicit about the MOPD pipeline stage names

Do not:

  • invent unpublished settings, dataset sizes, or hyperparameters
  • treat MOPD as ordinary RLHF
  • cite a benchmark number without saying which variant (base / BF16 / NVFP4) it belongs to
  • imply public reproducibility that the repo does not yet provide

© NVIDIA-NeMo, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 20 other files in skills/nemotron-ultra of NVIDIA-NeMo/Nemotron.

  • SKILL.md
  • INDEX.md
  • context/index.toml
  • context/quick-reference.md
  • model-card.md
  • paper/_overview.md
  • paper/architecture.md
  • paper/data.md
  • paper/evaluation.md
  • paper/inference.md
  • paper/infrastructure.md
  • paper/mopd/mtp-reasoning.md
  • paper/mopd/overview.md
  • paper/mopd/teachers.md
  • paper/mopd/warmup-results.md
  • paper/pretraining.md
  • paper/quantization.md
  • paper/safety.md
  • … and 3 more

Open the folder on GitHubat commit ca8c409

Compare with similar skills

Nemotron Ultra next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Nemotron Ultra compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Nemotron Ultra this skillNVIDIA-NeMo/Nemotron2.1k—~1.8kAutomated safety check: PassApache-2.0
Spark Environment Setupwshobson/agents40k—~2kAutomated safety check: PassMIT
Tao Finetune ClipNVIDIA/skills3.6k—~4kAutomated safety check: NotesApache-2.0
Tao Finetune Video ClipNVIDIA/skills3.6k—~3.5kAutomated safety check: NotesApache-2.0
Matlab Use Visual Inspectionmatlab/matlab-agentic-toolkit1.1k—~3.1kAutomated safety check: PassCustom licence
Setup Workshop Nemoclawbrevdev/workshop-build-an-agent146—~5.2kAutomated safety check: PassApache-2.0

Similar skills

  • Spark Environment Setup

    wshobson/agents

    Set up a working ML training/inference environment on NVIDIA DGX Spark (GB10, aarch64, CUDA 13).

    40k GitHub stars~2k tokensUpdated 6 days ago
    AI & LLM EngineeringAuto-check passed

  • Tao Finetune Clip

    NVIDIA/skills

    Official

    CLIP vision-language model for image-text retrieval, zero-shot classification, embedding extraction, ONNX export, and TensorRT deployment.

    3.6k GitHub stars~4k tokensUpdated yesterday
    AI & LLM EngineeringAuto-check: notes

  • Tao Finetune Video Clip

    NVIDIA/skills

    Official

    InternVideo2-CLIP L14 (TAO videoclip) for video-text retrieval, zero-shot classification, embedding extraction, LoRA fine-tuning, ONNX export, and TensorRT deployment.

    3.6k GitHub stars~3.5k tokensUpdated yesterday
    AI & LLM EngineeringAuto-check: notes

  • Matlab Use Visual Inspection

    matlab/matlab-agentic-toolkit

    Build machine vision inspection systems with MATLAB Visual Inspection Toolbox.

    1.1k GitHub stars~3.1k tokensUpdated 2 days ago
    AI & LLM EngineeringAuto-check passed

  • Setup Workshop Nemoclaw

    brevdev/workshop-build-an-agent

    Set up the NVIDIA "Build an Agent" DevX workshop as a working JupyterLab environment from INSIDE a locked-down OpenShell/NemoClaw sandbox, and hand the user the token URL + access commands.

    146 GitHub stars~5.2k tokensUpdated 3 days ago
    AI & LLM EngineeringAuto-check passed

  • Gptq

    Orchestra-Research/AI-Research-SKILLs

    Post-training 4-bit quantization for LLMs with minimal accuracy loss.

    13k GitHub starsUsed in 2 repos~2.9k tokens
    AI & LLM EngineeringAuto-check passed

More from NVIDIA-NeMo/Nemotron

All 10 skills in this repo

  • Nemotron Add Model

    NVIDIA-NeMo/Nemotron

    Onboard a new model family (Nemotron or third-party) into skills/ — paper chunks, recipe summaries, context packs, and model card.

    2.1k GitHub stars~2.2k tokensUpdated 4 days ago
    Auto-check passed

  • Nemotron Add Pattern

    NVIDIA-NeMo/Nemotron

    Add a cross-cutting decision pattern under src/nemotron/steps/patterns/.

    2.1k GitHub stars~1.4k tokensUpdated 4 days ago
    Auto-check passed

  • Nemotron Add Step

    NVIDIA-NeMo/Nemotron

    Add a new step under src/nemotron/steps/<category/<stepid/ — manifest (step.toml), runner glue, configs, and per-step README.md.

    2.1k GitHub stars~1.7k tokensUpdated 4 days ago
    Auto-check passed

  • Nemotron Customizer Airgap

    NVIDIA-NeMo/Nemotron

    Prepare, validate, build, and use Nemotron Customizer airgap image bundles for offline clusters.

    2.1k GitHub stars~1.2k tokensUpdated 4 days ago
    Auto-check passed

  • Nemotron Nano3

    NVIDIA-NeMo/Nemotron

    Reference desk for Nemotron 3 Nano / Llama-Nemotron Nano 3 — architecture, training data, recipes, evaluation, quantization, deployment.

    2.1k GitHub stars~1.9k tokensUpdated 4 days ago
    Auto-check passed

  • Nemotron Super3

    NVIDIA-NeMo/Nemotron

    Reference desk for NVIDIA Nemotron 3 Super — architecture, training data, recipes (pretrain/SFT/RL/eval/quantization), and deployment notes.

    2.1k GitHub stars~2.4k tokensUpdated 4 days ago
    Auto-check passed

Works with

Categories

Questions about Nemotron Ultra

What does Nemotron Ultra do?

Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference. Nemotron Ultra is an agent skill from NVIDIA-NeMo/Nemotron. Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference.

When should I use Nemotron Ultra?

Nemotron Ultra fits situations like: the user asks facts about Ultra rather than building a pipeline; tasks that involve LLM inference and serving; tasks that involve Fine-tuning.

How do I install Nemotron Ultra in Claude Code?

Run `npx skills add NVIDIA-NeMo/Nemotron --skill nemotron-ultra -a claude-code`. Or copy the skill folder (skills/nemotron-ultra in NVIDIA-NeMo/Nemotron) into .claude/skills/nemotron-ultra in your project. Claude Code loads it when a task matches its description.

How do I install Nemotron Ultra in Codex?

Run `npx skills add NVIDIA-NeMo/Nemotron --skill nemotron-ultra -a codex`. Or copy the skill folder (skills/nemotron-ultra in NVIDIA-NeMo/Nemotron) into .agents/skills/nemotron-ultra in your project. Codex loads it when a task matches its description.

Can I use Nemotron Ultra in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add NVIDIA-NeMo/Nemotron --skill nemotron-ultra -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/nemotron-ultra, .gemini/skills/nemotron-ultra, .github/skills/nemotron-ultra and .opencode/skills/nemotron-ultra in your project.

What does Nemotron Ultra need to run?

SKILL.md names no scripts, command-line tools or credentials: Nemotron Ultra is instructions for the agent only.

Does Nemotron Ultra access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Nemotron Ultra safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Nemotron Ultra use?

Nemotron Ultra is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Nemotron Ultra use?

About 1.8k tokens (SKILL.md is roughly 7.4k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Nemotron Ultra?

Skills that share tags, products or a category with Nemotron Ultra: Spark Environment Setup (wshobson/agents, 40k stars), Tao Finetune Clip (NVIDIA/skills, 3.6k stars), Tao Finetune Video Clip (NVIDIA/skills, 3.6k stars) and Matlab Use Visual Inspection (matlab/matlab-agentic-toolkit, 1.1k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Nemotron Ultra?

NVIDIA-NeMo (a GitHub organization) maintains it in NVIDIA-NeMo/Nemotron, which has 2,142 GitHub stars. The repository holds 10 skills in this directory. The repository was last updated on October 6, 2026.

Source: NVIDIA-NeMo/Nemotron on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.