# NVIDIA Nemotron Developer Repository
**Open and efficient models for agentic AI.** Training recipes, deployment guides, and use-case examples for the Nemotron family.
[](https://www.python.org/downloads/)
[](https://opensource.org/licenses/Apache-2.0)
[](CONTRIBUTING.md)
[](https://docs.nvidia.com/nemotron/latest/)
<div align="center">
[](https://www.youtube.com/watch?v=_y9SEtn1lU8)
**[Watch: Nemotron Overview](https://www.youtube.com/watch?v=_y9SEtn1lU8)**
</div>
---
> 🎉**Nemotron 3.5 Super** is now released — a 120B-A12B hybrid Mamba-Transformer LatentMoE vision-language model with native Multi-Token Prediction and up to 1M context, built for multimodal agentic reasoning over text, images, and video. See the [release blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/), the [technical report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Super-and-Lightning-Technical-Report.pdf), the [training recipe](./docs/nemotron/super35/README.md) and the [model weights](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16).
>
> 🎉**Nemotron 3.5 Lightning** is now released — a 30B-A3B hybrid Mamba-Transformer MoE with Multi-Token Prediction, built for the high-volume execution layer of long-running agents. See the [release blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/), the [training recipe](./docs/nemotron/lightning35/README.md), and the [model weights](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16).
>
> 🎉Nemotron 3 Ultra was [announced](https://www.youtube.com/live/q_umfWm8J28?t=4568s) at GTC San Jose 2026\. The model is open-source on [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), and the [training recipe](./docs/nemotron/ultra3/README.md) is now available in this repo\. To learn more, [see the usage guide](./usage-cookbook/Nemotron-3-Ultra/README.md)\!
>
> 🎉**Nemotron 3 Nano Omni** is now released — a 30B-A3B hybrid Mamba-Transformer MoE with native text, image, video, and audio support, designed as a multimodal perception sub-agent for agentic AI. See the [release blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-nano-omni-powers-multimodal-agent-reasoning-in-a-single-efficient-open-model/), the [training recipe](./docs/nemotron/omni3/README.md), and the [model weights](https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16).
---
## Why Nemotron?
| | |
|---|---|
| **Open Models** | Fully transparent training data, techniques, and weights for community innovation |
| **Compute Efficiency** | Model pruning and optimization enabling higher throughput via TensorRT-LLM |
| **High Accuracy** | Built on frontier open models with human-aligned reasoning for agentic workflows |
| **Flexible Deployment** | Deploy anywhere: edge, single GPU, or data center with NIM microservices |
---
## Use from Claude Code
This repo ships a Claude Code plugin called **`nemotron-customize`** that turns the step catalog under [`src/nemotron/steps/`](./src/nemotron/steps/) into a guided, repo-native pipeline builder.
Install once:
```text
/plugin marketplace add NVIDIA-NeMo/Nemotron
/plugin install nemotron-customize@nvidia-nemotron
```
Then, **start Claude Code from the repo root** and invoke the skill:
```bash
cd /path/to/Nemotron # repo root: must contain pyproject.toml and src/nemotron/steps/
claude
```
```text
/nemotron-customize
```
The skill resolves all file paths against your current working directory, so it must be invoked from the Nemotron checkout root. Running it from a subdirectory will cause file reads to fail.
The skill plans the step DAG, validates artifact wiring, and emits the YAML configs needed to run the requested pipeline. See [`skills/nemotron-customize/SKILL.md`](./skills/nemotron-customize/SKILL.md) for the full contract.
> The marketplace installs **only** `nemotron-customize`. The other folders under [`skills/`](./skills/) (model knowledge bases, contributor add-`*` skills) stay on disk for repo browsing but are not loaded as plugins.
---
## Repository Overview
```
nemotron/
│
├── src/nemotron/steps/ Modular building blocks for training, eval, SDG, and more
│
├── src/nemotron/recipes/ Training recipes (complete, reproducible pipelines)
│
├── usage-cookbook/ Usage cookbooks (deployment and model usage guides)
│
└── use-case-examples/ Examples of leveraging Nemotron in agentic workflows
```
### Which section should I use?
| | **Nemotron Steps** | **Training Recipes** | **Usage Cookbooks** | **Use Case Examples** |
|---|---|---|---|---|
| **Purpose** | Full lifecycle building blocks, chain data prep, training, eval and other steps | Reproduce full training pipelines from raw data to model | Deploy and use trained models | Build end-to-end applications |
| **Format** | The `nemotron steps` CLI and YAML configs | Python packages with configs, scripts, and evaluation | Jupyter notebooks with step-by-step guides | Jupyter notebooks and scripts |
| **When to use** | You want to run one stage in isolation or compose a custom pipeline | You want to train, fine-tune, or understand how a model was built | You have a model and want to deploy or run inference | You want to build an application (RAG, agents, tool use) |
| **Location** | [`src/nemotron/steps/`](./src/nemotron/steps/) | [`src/nemotron/recipes/`](./src/nemotron/recipes/) | [`usage-cookbook/`](./usage-cookbook/) | [`use-case-examples/`](./use-case-examples/) |
---
## What is Nemotron?
[NVIDIA Nemotron](https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/) is a family of open, high-efficiency multimodal models purpose-built for agentic AI.
**Model Tiers:**
- **Nano** — Optimized for edge and PC deployments
- **Lightning** — Fast, high-volume task execution for long-running agents
- **Super** — Single GPU deployment with highest throughput
- **Ultra** — Multi-GPU datacenter applications
Nemotron models excel at coding, math, scientific reasoning, tool calling, instruction following, and visual reasoning. Deploy across edge, single GPU, or data center environments with support for NeMo, TensorRT-LLM, vLLM, SGLang, and NIM microservices.
---
## Nemotron Steps
A *Nemotron step* is a named, reusable unit of work that you invoke with the `nemotron steps` CLI.
Each step packages a description of the work it performs, the artifacts it consumes and produces, and one or more named configurations that supply parameter values.
Steps live under [`src/nemotron/steps/`](./src/nemotron/steps/), and the CLI discovers them at startup.
The training recipes in the next section are composed from these steps.
Run a step on its own when you want one stage, or chain steps together when you need a different pipeline shape than the published recipes.
### Step Categories
The catalog covers the full training lifecycle.
- Data curation and preparation with `curate/*` and `data_prep/*`.
- Synthetic data generation (SDG) with `sdg/*`.
This includes the [Persona MCQ pipeline](src/nemotron/steps/sdg/persona_mcq/README.md),
available as `nemotron steps run sdg/persona_mcq`.
- Corpus translation with `translate/*`.
- Bring-your-own benchmark generation with `byob/*`.
- Pretraining, supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), and reinforcement learning (RL) with `pretrain/*`, `sft/*`, `peft/*`, and `rl/*`.
- Checkpoint conversion and model optimization with `convert/*` and `optimize/*`.
- Benchmark evaluation with `eval/*`.
- Execution-profile setup with `env/*`.
### Documentation
- [About Nemotron Steps](docs/steps/index.md) is the entry point for the step model.
- [Nemotron Steps Basics](docs/steps/basics.md) defines *step*, *configuration*, *environment profile*, and *artifact*.
- [Getting Started With Steps](docs/steps/getting-started.md) walks through the CLI on a small example.
---
## Training Recipes
The Nemotron repository provides reproducible training pipelines from raw data to deployment-ready models. These implementations reflect how large language models are actually trained: careful experimentation, validation gates, and systematic optimization.
### Why Complete Pipelines?
Training a production model involves interconnected components. Isolated examples miss how stages interact. Complete pipelines show:
- **How data quality affects downstream performance** across pretraining, SFT, and RL
- **Which training techniques actually work together**, not just in theory
- **Where validation gates prevent failures** and maintain reproducibility
- **How to balance competing objectives** across stages
Because these are complete systems, you can extract specific techniques with confidence. Each component has been proven to work in context.
### Each Recipe Includes
- 🎨 **Synthetic Data Generation** - Scripts to generate synthetic datasets using [NVIDIA-NeMo/DataDesigner](https://github.com/NVIDIA-NeMo/DataDesigner)
- 🗂️ **Data Curation** - Scripts to prepare training data using [NVIDIA NeMo Curator](https://github.com/NVIDIA/NeMo-Curator) for scalable data processing, filtering, and quality enhancement
- 🔁 **Training** - Complete training loops with hyperparameters using:
- [NVIDIA-NeMo/Megatron-Bridge](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main) for Megatron models
- [NVIDIA-NeMo/Automodel](https://github.com/NVIDIA-NeMo/Automodel) for HuggingFace models
- [NVIDIA-NeMo/NeMo-RL](https://github.com/NVIDIA-NeMo/RL/tree/main) when RL is needed
- Includes GPU-accelerated last-mile data processing (tokenization + optional sequence packing) for optimal training efficiency
- 📊 **Evaluation** - Benchmark evaluation on standard suites using [NVIDIA NeMo Evaluator](https://github.com/NVIDIA-NeMo/Evaluator)
- 📖 **Documentation** - Detailed explanations of each stage
### Available Recipes and Workflow Maps
Where a model's canonical executable pipeline is maintained in another NVIDIA
repository, the guide is an integration map rather than a fork of that code.
Nemotron 3.5 Transcribe follows this pattern with NeMo Speech.
| Model | Description | Stages | Guide |
|-------|-------------|--------|-------|
| **[Nemotron 3 Ultra](docs/nemotron/ultra3/README.md)** | 550B total / 55B active hybrid Mamba-Attention LatentMoE Transformer with MTP and 1M context — NVIDIA's largest Nemotron 3 model for datacenter-scale agentic reasoning | Pretrain → SFT → RLVR → MOPD | [Training Guide](docs/nemotron/ultra3/README.md) |
| **[Nemotron 3 Super](docs/nemotron/super3/README.md)** | 120.6B total / 12.7B active Hybrid Mamba Latent MoE Transformer for frontier reasoning, coding, and agentic tasks | Pretrain → SFT → RL | [Training Guide](docs/nemotron/super3/README.md) |
| **[Nemotron 3.5 Super](docs/nemotron/super35/README.md)** | 120B total / 12B active hybrid Mamba-Transformer LatentMoE vision-language model with native MTP — text, image, and video reasoning for agentic workloads | Pretrain → SFT → RLVR → Teachers → MOPD → Eval | [Training Guide](docs/nemotron/super35/README.md) |
| **[Nemotron 3 Nano](docs/nemotron/nano3/README.md)** | 31.6B total / 3.6B active MoE Hybrid Mamba-Transformer for agentic reasoning | Pretrain → SFT → RL | [Training Guide](docs/nemotron/nano3/README.md) |
| **[Nemotron 3.5 Lightning](docs/nemotron/lightning35/README.md)** | 30B total / 3B active hybrid Mamba-Transformer MoE with Multi-Token Prediction | Pretrain → SFT → RL → Quantization | [Training Guide](docs/nemotron/lightning35/README.md) |
| **[Nemotron 3.5 Transcribe](docs/nemotron/transcribe35/README.md)** | 34B total / 4.3B active offline speech-and-audio model based on Nemotron 3.5 Lightning | Main-stage SALM construction or released-checkpoint adaptation → target branches (optional DFlash2 draft; optional quantization) → Inference | [Recipe Map](docs/nemotron/transcribe35/README.md) |
| **[Nemotron 3 Nano Omni](docs/nemotron/omni3/README.md)** | 30B total / 3B active hybrid Mamba-Transformer MoE — native text, image, video, and audio for agentic multimodal perception | SFT → RL (MPO / text / vision) → Eval | [Training Guide](docs/nemotron/omni3/README.md) |
| **[Nemotron Embed](docs/nemotron/embed/index.md)** | Fine-tune text and multimodal embedding models for domain-specific retrieval | SDG → Data Prep → Finetune → Eval → Export (Llama only) → Deploy | [Training Guide](docs/nemotron/embed/index.md) |
| **[Nemotron Rerank](docs/nemotron/rerank/index.md)** | Fine-tune text and multimodal rerankers to score and reorder retrieved documents | SDG → Data Prep → Finetune → Eval → Export (Llama only) → Deploy | [Training Guide](docs/nemotron/rerank/index.md) |
### Nemotron 3 Ultra
A training recipe for NVIDIA's largest Nemotron 3 model — a 550B-A55B hybrid Mamba-Attention Mixture-of-Experts Transformer with LatentMoE and multi-token prediction (MTP), pretrained in NVFP4 and extended to 1M-token context for datacenter-scale agentic reasoning.
> **Open-Source Data Only**: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.
**Model Specifications**:
- 550B total / 55B active parameters (MoE)
- Hybrid Mamba-Attention architecture with LatentMoE + two shared-weight MTP layers
- 20T pretraining tokens in NVFP4, two-phase data curriculum
- Up to 1M (1,048,576) context length
- Full program: Pretrain → SFT → RLVR → MOPD → MTP Boosting (this recipe covers **Pretrain → SFT**)
**What You Can Extract**:
- Two-phase pretraining data mixture (tech-report Figure 4) over the open Nemotron datasets
- Ray-based data prep: tokenize raw datasets → Megatron bin/idx (pretrain) and pack chat data → Parquet (SFT)
- New open pretraining datasets: Specialized-v1.2 (Multiple-Choice / Generative / Fact-Seeking / Moral-Scenarios) and Legal-v1
- Stage-local container builds (Day-0 Megatron-Bridge) for both pretrain and SFT
- Megatron-Bridge training at Ultra scale (TP=2 / PP=12 / EP=32 pretrain, PP=6 SFT)
**Resources**:
- [Training Guide](docs/nemotron/ultra3/README.md)
- [Usage Guide](./usage-cookbook/Nemotron-3-Ultra/README.md)
- [Model Weights (BF16)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
### Nemotron 3 Super
A complete training recipe for the frontier Hybrid Mamba Latent Mixture-of-Experts Transformer model with state-of-the-art reasoning, coding, and agentic capabilities.
> **Open-Source Data Only**: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.
**Model Specifications**:
- 120B total / 12B active parameters
- Multi-stage RL pipeline: 3× RLVR + 2× SWE-RL + RLHF across 21 reward environments
- Asynchronous GRPO with decoupled training and inference
**What You Can Extract**:
- Large-scale pretraining with data curriculum
- Multi-domain SFT pipeline
- Multi-environment RLVR with 21 simultaneous reward environments
- SWE-RL with container-isolated sandbox execution
- GenRM-based RLHF with principle-following rewards
- Asynchronous GRPO at 1K GPU scale
**Resources**:
- [Training Guide](docs/nemotron/super3/README.md)
- [Tech Report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf)
- [Model Weights (Instruct)](https://huggingface.co/nvidia/Nemotron-3-Super-49B-v1)
### Nemotron 3.5 Super
A training recipe for Nemotron 3.5 Super, a 120B-A12B hybrid Mamba-Transformer LatentMoE vision-language model. SFT turns its text-only base into a vision-language model; RLVR, six specialist teachers, and two-phase multi-teacher on-policy distillation (MOPD) then post-train it into a multimodal agent.
> **Pending Open-Source Release**: The SFT and RL data used for Super 3.5 are not yet published, and the agentic teacher's environment is not in public NeMo Gym. The configs, CLI, and RL stack (NeMo-RL `super-v3.5-posttraining`) are public; see the [training guide](docs/nemotron/super35/README.md#open-source-status) for what has been verified.
**Model Specifications**:
- 120B total / 12B active parameters, initialized from Nemotron 3 Super
- C-RADIO vision encoder + temporal video embedder (text, image, and video; no audio)
- Single recursively shared MTP head for speculative decoding
- Up to 1M context
- Released as Base BF16, post-trained BF16, and NVFP4 checkpoints
**What You Can Extract**:
- Megatron-Bridge vision-language pretraining, two-stage SFT, PEFT, and freeze controls for the language model, vision encoder, and projector
- Multi-environment RLVR with NeMo Gym, including competitive-programming and finance-agent environments
- Specialist teachers for reasoning, vision tool use, video reasoning and tracking, vision reasoning, and alignment (GenRM-based RLHF)
- Two-phase MOPD with per-domain teacher routing, a self-anchor, and a proximal-teacher target
- NVFP4 PTQ + QAD reaching 99.52% median accuracy recovery (not a CLI stage; the released precision recipe runs through Model Optimizer's W4A4 tutorial)
**Resources**:
- [Release Blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/)
- [Training Guide](docs/nemotron/super35/README.md)
- [Tech Report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Super-and-Lightning-Technical-Report.pdf)
- [NeMo-RL Super 3.5 Guide](https://github.com/NVIDIA-NeMo/RL/blob/super-v3.5-posttraining/docs/guides/models/nemotron/nemotron-3.5-super-vl.md)
- [NeMo Gym Evaluation Reproducibility Guide](https://github.com/NVIDIA-NeMo/Gym/blob/jakub/marlin/nemotron_recipes/super-3.5/reproducibility.md)
- [Quantization Guide (PTQ + QAD)](docs/nemotron/super35/quantization.md)
- [Model Weights (Base)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16)
- [Model Weights (Instruct)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16)
- [Model Weights (NVFP4)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4)
### Nemotron 3.5 Lightning
A complete training recipe for the 30B-A3B hybrid Mamba-Transformer MoE with Multi-Token Prediction, built for fast, accurate specialized task execution in long-running agents. Weights, data, and recipes are released under OpenMDW-1.1.
> **Open-Source Data Only**: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the published benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.
**Model Specifications**:
- 30B total parameters, 3B active per forward pass
- Hybrid Mamba-Transformer with sparse MoE and Multi-Token Prediction (MTP)
- Ships with DSpark and DFlash draft models for speculative decoding, plus an NVFP4 checkpoint
- Optimized for the high-volume execution layer of always-on agents: tool calls, result validation, and subagent delegation
**What You Can Extract**:
- Pretraining with MTP heads (repeated-layer MTP, 0.3 loss scaling)
- Packed-sequence SFT with the Lightning chat template
- Multi-environment RLVR with GRPO via NeMo Gym, MTP heads kept training
- Gym-native benchmark evaluation (GPQA, SciCode, and the published reference suite)
- NVFP4 quantization with four-over-six PTQ + Quantization-Aware Distillation (QAD)
- File-based (manifest) or W&B artifact lineage across all stages
**Resources**:
- [Release Blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/)
- [Training Guide](docs/nemotron/lightning35/README.md)
- [Tech Report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Super-and-Lightning-Technical-Report.pdf)
- [NeMo-RL Lightning Guide](https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/nemotron-3.5-lightning.md)
- [DGX Station Customization Recipes](./usage-cookbook/Nemotron-3.5-Lightning/dgx-station-recipes/README.md)
- [Quantization Guide (PTQ + QAD)](docs/nemotron/lightning35/quantization.md)
- [Model Weights (Base)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16)
- [Model Weights (Instruct)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16)
- [Model Weights (NVFP4)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4)
### Nemotron 3.5 Transcribe
An offline speech-and-audio language model for instruction-controlled multilingual transcription, contextual ASR, speaker-attributed and rich transcription, timestamps, speech translation, language identification, and audio understanding.
The canonical executable workflows live in [NeMo Speech](https://github.com/NVIDIA-NeMo/Speech). The Nemotron contribution provides a model-specific recipe map, reference skill, and usage cookbook without copying the NeMo Speech main-stage training, fine-tuning, DFlash2, or quantization procedures.
**Model Specifications**:
- Approximately 34B total / 4.3B active parameters, including a parallel ASR and speaker encoder
- Nemotron 3.5 Lightning hybrid Mamba-Transformer MoE language backbone
- 40 transcription languages, up to one-hour recordings and 65,536-token training context; audio understanding was evaluated only in English
- Native MTP, optional DFlash2 draft, and optional CTC word timestamps
- BF16, FP8, and NVFP4 checkpoints under OpenMDW-1.1
**Available Public Workflows**:
- Main-stage SALM construction from pretrained language and speech components
- Downstream checkpoint fine-tuning and task adaptation with NeMo Speech
- DFlash/DFlash2 draft training with a frozen speech target
- FP8 and NVFP4 quantization, including the audio encoder and MTP head
- vLLM serving through the NeMo Speech plugin
The public main-stage configuration is a parameterized construction template,
not the exact released-checkpoint reproduction setup. Production data blends,
component checkpoints, cluster values, RL post-training, timestamp-head
training, and reproducible evaluation assets are not included; the
[recipe map](docs/nemotron/transcribe35/README.md#open-recipe-coverage) tracks
those gaps.
**Resources**:
- [Recipe Map](docs/nemotron/transcribe35/README.md)
- [Usage Cookbook](./usage-cookbook/Nemotron-3.5-Transcribe/README.md)
- [Reference Skill](skills/nemotron-transcribe/SKILL.md)
- [Technical Report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Transcribe-Technical-Report.pdf)
- [Model Weights (BF16)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B)
- [Model Weights (FP8)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-FP8)
- [Model Weights (NVFP4)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-NVFP4)
- [Draft Model (DFlash2)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-DFlash2)
### Nemotron 3 Nano
A complete training recipe for the open, efficient Mixture-of-Experts hybrid Mamba-Transformer model optimized for agentic reasoning.
> **Open-Source Data Only**: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.
**Model Specifications**:
- 31.6B total parameters, 3.6B active per forward pass
- 25 trillion pretraining tokens with curriculum learning
- Up to 1M context length
- 3.3x higher inference throughput than similarly sized models
**What You Can Extract**:
- Curriculum-based pretraining with two-phase data mixture
- Long-context extension via CPT methodology
- Multi-domain SFT with 12+ data sources
- InfinityByte cross-domain code synthesis
- Tool-calling fine-tuning and budget-controlled reasoning
- Multi-environment RLVR with GRPO
- GenRM reward modeling with circular comparison
- DPO for tool hallucination reduction
**Resources**:
- [Training Guide](docs/nemotron/nano3/README.md)
- [Tech Report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Nano-Technical-Report.pdf)
- [Model Weights (Base)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16)
- [Model Weights (Instruct)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16)
- [Model Weights (FP8)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8)
### Nemotron 3 Nano Omni
A multimodal training recipe for the 30B-A3B hybrid Mamba-Transformer Mixture-of-Experts model. Native support for text, image, video, and audio in a single decoder, designed as a perception sub-agent for agentic AI.
> **Open-Source Data Only**: These recipes train exclusively on the open-sourced subset of training data (e.g., CORD-v2 for SFT, public MMPR / MMPR-Tiny for RL). Results will differ from the release benchmarks, which used additional internal datasets. Use these recipes as reference implementations to apply the methodology with your own data.
**Model Specifications**:
- 30B total / 3B active parameters (A3B MoE)
- Hybrid architecture: Mamba layers (sequence/memory efficiency) + transformer layers (reasoning), with a unified text decoder
- Native modalities: text, image, video, audio
- Vision encoder: C-RADIOv4-H · Audio encoder: NVIDIA Parakeet · Video pipeline: 3D convolutions + Efficient Video Sampling (EVS)
- Context length: progressively scaled 16K → 49K → 262K
- Best-in-class on MMlongbench-Doc, OCRBenchV2; leading on WorldSense, DailyOmni, VoiceBench
- Up to ~9.2× greater video-reasoning system capacity, ~7.4× on multi-document workloads vs. comparable open omni models
- License: NVIDIA Nemotron Open Model License (enterprise-friendly, on-prem and any deployment)
**What You Can Extract**:
- Multimodal SFT pipeline using Megatron-Bridge with the Valor32k recipe family (open-dataset CORD-v2 default + Valor32k variants)
- Progressive context scaling: 16K → 49K → 262K
- Multimodal preference optimization (MPO) on the public MMPR dataset
- Text-only GRPO continuation of alignment via NeMo-RL
- Vision GRPO on MMPR-Tiny
- Inline NVIDIA stack: Megatron-Bridge for SFT, NeMo-RL (`nano-v3-omni` branch with the omni vllm fork as a submodule) for RL
- Cookbook-style end-to-end recipe (build → data prep → SFT → RL → eval) reproducing the release training stages
**Resources**:
- [Training Guide](docs/nemotron/omni3/README.md)
- [Release Blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-nano-omni-powers-multimodal-agent-reasoning-in-a-single-efficient-open-model/)
- [Model Weights (BF16)](https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16)
- [Image Training Data](https://huggingface.co/datasets/nvidia/Nemotron-Image-Training-v3)
- Upstream pre-training recipe: [`NVIDIA-NeMo/Megatron-Bridge` `nemotron_3_omni`](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/nemotron_3_omni)
- Upstream RL recipe: [`NVIDIA-NeMo/RL` `nano-v3-omni`](https://github.com/NVIDIA-NeMo/RL/tree/nano-v3-omni)
---
## Usage Cookbooks
Practical deployment and model usage guides for Nemotron models.
| Model | Best For | Key Features | Resources |
|-------|----------|--------------|-----------|
| [**Nemotron 3 Ultra 550B A55B**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) | Long-running coding, research, and enterprise agentic workflows | 1M context, 550B/55B MoE, MTP, multi-GPU deployment, agent harness configs | [Cookbooks](./usage-cookbook/Nemotron-3-Ultra) |
| [**Nemotron 3 Super 120B A12B**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16) | Production deployments needing strong reasoning | 1M context, in NVFP4 single B200, RAG & tool calling | [Cookbooks](./usage-cookbook/Nemotron-3-Super) |
| [**Nemotron 3 Nano 30B A3B**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) | Resource-constrained environments | 1M context, sparse MoE hybrid Mamba-2, controllable reasoning | [Cookbooks](./usage-cookbook/Nemotron-3-Nano) |
| [**Nemotron 3.5 Transcribe 30B A4B**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) | Multilingual transcription, translation, speaker attribution, timestamps, and English-only evaluated audio understanding | 40 transcription languages, context prompting, MTP, DFlash2 draft, BF16/FP8/NVFP4 targets | [Cookbook](./usage-cookbook/Nemotron-3.5-Transcribe) |
| [**Llama-3.1-Nemotron-Nano-8B-v1**](https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-8B-v1) | Small-footprint OCI deployments | Validated on private OKE in Phoenix with `vLLM`, OCI Bastion service, tool calling, and OpenAI-compatible `/v1` inference; provides a reproducible OCI path comparable to common AWS GPU/Kubernetes deployment patterns | [Cookbooks](./usage-cookbook/Llama-3.1-Nemotron-Nano-8B-v1) |
| [**NVIDIA-Nemotron-Nano-12B-v2-VL**](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL) | Document intelligence and video understanding | 12B VLM, video reasoning, Efficient Video Sampling | [Cookbooks](./usage-cookbook/Nemotron-Nano2-VL/) |
| [**Llama-3.1-Nemotron-Safety-Guard-8B-v3**](https://huggingface.co/nvidia/Llama-3.1-Nemotron-Safety-Guard-8B-v3) | Multilingual content moderation | 9 languages, 23 safety categories | [Cookbooks](./usage-cookbook/Llama-3.1-Nemotron-Safety-Guard-V3/) |
| **Nemotron-Parse** | Document parsing for RAG and AI agents | Table extraction, semantic segmentation | [Cookbooks](./usage-cookbook/Nemotron-Parse-v1.1/) |
---
## Use Case Examples
End-to-end examples demonstrating practical applications in the [`use-case-examples/`](./use-case-examples/) directory:
- **Agentic Workflows** — Multi-step AI agents with planning, context management, and external tools
- **RAG Systems** — Pipelines combining retrieval with Nemotron models for grounded outputs
- **Tool Integration** — Structured tool calling, function execution, and data enrichment
- **Production Patterns** — Scalability, monitoring, and deployment architectures
---
## Nemotron Open Datasets
More than just weights, recipes, and libraries: Nemotron is committed to opening data across many domains, training phases, and use cases.
<details>
<summary><strong>Nemotron Data Catalogue</strong></summary>
*A comprehensive collection of NVIDIA Nemotron datasets spanning pre-training, post-training, reinforcement learning, multimodal, safety, and domain-specific applications. These openly available datasets power the Nemotron family of models for agentic AI development.*
---
<details>
<summary><strong>Code</strong></summary>
*Datasets for training code generation, competitive programming, and software engineering capabilities across multiple programming languages.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-CC-Code-v1](https://huggingface.co/datasets/nvidia/Nemotron-CC-Code-v1) | Pre-training | NVIDIA Data Agreement | Nemotron 3 Nano | 427.9B tokens from Common Crawl code pages using Lynx + LLM pipeline |
| [Nemotron-Pretraining-Code-v1](https://huggingface.co/datasets/nvidia/Nemotron-Pretraining-Code-v1) | Pre-training | NVIDIA Data Agreement | Nemotron Nano 2 | GitHub-sourced code corpus for Nemotron Nano 2 |
| [Nemotron-Pretraining-Code-v2](https://huggingface.co/datasets/nvidia/Nemotron-Pretraining-Code-v2) | Pre-training | NVIDIA Data Agreement | Nemotron 3 Nano | Updated GitHub code + synthetic QA with STEM reasoning |
| [Nemotron-Cascade-RL-SWE](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-SWE) | RL Training | CC-BY-4.0 | Nemotron 3 | SWE code repair from SWE-Bench, SWE-Smith, R2E-Gym |
| [Nemotron-Competitive-Programming-v1](https://huggingface.co/datasets/nvidia/Nemotron-Competitive-Programming-v1) | SFT | CC-BY-4.0 | Nemotron 3 | 2M+ Python and 1M+ C++ samples across 34K competitive programming questions |
| [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning) | SFT | CC-BY-4.0 | OpenCode-Nemotron | 735K Python samples across 28K competitive programming questions |
| [OpenCodeReasoning-2](https://huggingface.co/datasets/nvidia/OpenCodeReasoning-2) | SFT | CC-BY-4.0 | OpenCode-Nemotron | 2.5M samples (1.4M Python, 1.1M C++) with code completion and critique |
| [Scoring-Verifiers](https://huggingface.co/datasets/nvidia/Scoring-Verifiers) | Evaluation | CC-BY-4.0 | — | Benchmark for test case generation and code reward models |
</details>
---
<details>
<summary><strong>Math</strong></summary>
*Mathematical reasoning datasets ranging from pre-training corpora to advanced problem-solving with chain-of-thought and tool-integrated reasoning. Includes the AIMO-2 competition winning dataset.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-CC-Math-v1](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1) | Pre-training | NVIDIA Data Agreement | Nemotron Nano 2, Nemotron 3 Nano | 133B-token math dataset from Common Crawl using Lynx + LLM pipeline |
| [Nemotron-Math-Proofs-v1](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1) | SFT | CC-BY-4.0 | Nemotron 3 Nano | Mathematical proofs dataset for Nemotron 3 post-training |
| [Nemotron-Math-v2](https://huggingface.co/datasets/nvidia/Nemotron-Math-v2) | SFT | CC-BY-4.0 | Nemotron 3 | 347K samples and 7M reasoning trajectories for Deeper Math Reasoning |
| [Nemotron-CrossThink](https://huggingface.co/datasets/nvidia/Nemotron-CrossThink) | RL Training | CC-BY-4.0 | Nemotron 3 | Multi-domain QA with MCQ and open-ended formats for verifiable rewards |
| [OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning) | SFT | CC-BY-4.0 | OpenMath-Nemotron | 5.68M samples, 306K problems from AoPS with CoT/TIR (AIMO-2 winner) |
</details>
---
<details>
<summary><strong>Science / STEM</strong></summary>
*Scientific reasoning datasets covering chemistry, physics, and general STEM domains for training models on scientific question answering and reasoning.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-Science-v1](https://huggingface.co/datasets/nvidia/Nemotron-Science-v1) | SFT | CC-BY-4.0 | Nemotron 3 Nano | Synthetic science reasoning (MCQA + chemistry RQA) |
</details>
---
<details>
<summary><strong>General / Web</strong></summary>
*Large-scale web-crawled and curated datasets for pre-training and post-training, including multilingual data and general instruction-following capabilities.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-CC-v2.1](https://huggingface.co/datasets/nvidia/Nemotron-CC-v2.1) | Pre-training | NVIDIA Data Agreement | Nemotron 3 Nano | 2.5T tokens English web data with synthetic rephrases and translations |
| [Nemotron-CC-v2](https://huggingface.co/datasets/nvidia/Nemotron-CC-v2) | Pre-training | NVIDIA Data Agreement | Nemotron Nano 2 | 6.6T tokens quality-filtered Common Crawl with multilingual Q&A |
| [Nemotron-Pretraining-Dataset-sample](https://huggingface.co/datasets/nvidia/Nemotron-Pretraining-Dataset-sample) | Pre-training (Sample) | NVIDIA Data Agreement | — | Sample subset of Nemotron pre-training corpus for experimentation |
| [Llama-Nemotron-Post-Training-Dataset](https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset) | SFT + RL | CC-BY-4.0 | Llama-Nemotron Ultra/Super/Nano | Math, code, reasoning data (2.2M math, 500K code) |
| [Nemotron-Post-Training-Dataset-v1](https://huggingface.co/datasets/nvidia/Nemotron-Post-Training-Dataset-v1) | SFT | CC-BY-4.0 | Llama-3.3-Nemotron-Super-49B-v1.5 | Math, code, STEM, tool calling |
| [Nemotron-Post-Training-Dataset-v2](https://huggingface.co/datasets/nvidia/Nemotron-Post-Training-Dataset-v2) | SFT + RL | CC-BY-4.0 | Llama-Nemotron | Multilingual extension (Spanish, French, German, Italian, Japanese) |
| [Nemotron-3-Nano-RL-Training-Blend](https://huggingface.co/datasets/nvidia/Nemotron-3-Nano-RL-Training-Blend) | RL Training | CC-BY-4.0 | Nemotron-3-Nano-30B-A3B | Curated multi-domain blend for Nemotron 3 Nano |
| [Nemotron-RL-knowledge-web_search-mcqa](https://huggingface.co/datasets/nvidia/Nemotron-RL-knowledge-web_search-mcqa) | RL Training | ODC-BY-1.0 | Nemotron 3 | Web search and multiple-choice QA tasks for NeMo Gym |
</details>
---
<details>
<summary><strong>Chat / Instruction Following</strong></summary>
*Datasets for training conversational AI with strong instruction-following capabilities, structured output generation, and multi-turn dialogue.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-Instruction-Following-Chat-v1](https://huggingface.co/datasets/nvidia/Nemotron-Instruction-Following-Chat-v1) | SFT | CC-BY-4.0 | Nemotron 3 Nano | Multi-turn chat and structured output generation |
| [Nemotron-RL-instruction_following](https://huggingface.co/datasets/nvidia/Nemotron-RL-instruction_following) | RL Training | ODC-BY-1.0 | Nemotron 3 | Verifiable instruction adherence from WildChat-1M + Open-Instruct |
| [Nemotron-RL-instruction_following-structured_outputs](https://huggingface.co/datasets/nvidia/Nemotron-RL-instruction_following-structured_outputs) | RL Training | ODC-BY-1.0 | Nemotron 3 | JSON schema-constrained output formatting tests |
| [Nemotron-Cascade-RL-Instruction-Following](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-Instruction-Following) | RL Training | ODC-BY-1.0 | Nemotron 3 | 108K samples for instruction-following RL |
</details>
---
<details>
<summary><strong>Agentic / Tool Use</strong></summary>
*Datasets for training AI agents with tool calling, multi-step workflows, and agentic reasoning capabilities.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-Agentic-v1](https://huggingface.co/datasets/nvidia/Nemotron-Agentic-v1) | SFT | CC-BY-4.0 | Nemotron 3 Nano | Multi-turn trajectories for conversational tool use and agentic workflows |
| [Nemotron-RL-agent-workplace_assistant](https://huggingface.co/datasets/nvidia/Nemotron-RL-agent-workplace_assistant) | RL Training | ODC-BY-1.0 | Nemotron 3 | Workplace assistant agent tasks for NeMo Gym |
</details>
---
<details>
<summary><strong>Alignment / Reward Modeling</strong></summary>
*Human preference and reward modeling datasets for RLHF, SteerLM training, and model alignment. Powers top-performing reward models on RM-Bench and JudgeBench.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [HelpSteer3](https://huggingface.co/datasets/nvidia/HelpSteer3) | Reward Modeling | CC-BY-4.0 | Nemotron 3 Nano, Llama-Nemotron Super 49B | 40K+ samples; top on RM-Bench/JudgeBench with preference, feedback, edit-quality |
| [HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2) | Reward Modeling | CC-BY-4.0 | Nemotron-4-340B-Reward, Llama-3.1-Nemotron-70B-Reward | 21K samples with 5 attributes |
| [HelpSteer](https://huggingface.co/datasets/nvidia/HelpSteer) | SteerLM Training | CC-BY-4.0 | Nemotron-4 SteerLM | 37K samples (helpfulness, correctness, coherence, complexity, verbosity) |
| [Daring-Anteater](https://huggingface.co/datasets/nvidia/Daring-Anteater) | SFT/RLHF | CC-BY-4.0 | Nemotron-4-340B-Instruct | Instruction tuning dataset; synthetic subsets + FinQA, wikitablequestions |
| [sft_datablend_v1](https://huggingface.co/datasets/nvidia/sft_datablend_v1) | SFT | CC-BY-4.0 | — | SFT data blend for RLHF pipeline |
</details>
---
<details>
<summary><strong>Vision-Language / Multimodal</strong></summary>
*High-quality VLM training data for document intelligence, OCR, image reasoning, video QA, and chain-of-thought visual understanding.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-VLM-Dataset-v2](https://huggingface.co/datasets/nvidia/Nemotron-VLM-Dataset-v2) | VLM Training | CC-BY-4.0 (some CC-BY-SA-4.0) | Nemotron VLM | 8M samples for OCR, image reasoning, video QA with chain-of-thought |
| [Llama-Nemotron-VLM-Dataset-v1](https://huggingface.co/datasets/nvidia/Llama-Nemotron-VLM-Dataset-v1) | VLM Training | CC-BY-4.0 (some CC-BY-SA-4.0) | Llama-3.1-Nemotron-Nano-VL-8B | 3M samples for visual question answering and captioning |
</details>
---
<details>
<summary><strong>Physical AI / Robotics</strong></summary>
*Datasets for embodied reasoning, physical common sense, and robotic manipulation. Powers Cosmos-Reason1 for physical AI applications.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Cosmos-Reason1-SFT-Dataset](https://huggingface.co/datasets/nvidia/Cosmos-Reason1-SFT-Dataset) | SFT | CC-BY-4.0 | Cosmos-Reason1-7B | Video-text pairs for robotics, ego-centric demos, AV reasoning |
| [Cosmos-Reason1-RL-Dataset](https://huggingface.co/datasets/nvidia/Cosmos-Reason1-RL-Dataset) | RL Training | CC-BY-4.0 | Cosmos-Reason1-7B | RL data for physical common sense and embodied reasoning |
| [Cosmos-Reason1-Benchmark](https://huggingface.co/datasets/nvidia/Cosmos-Reason1-Benchmark) | Evaluation | CC-BY-4.0 | — | Benchmark for embodied reasoning (robotics, HoloAssist, AV) |
| [PhysicalAI-Robotics-Manipulation-Augmented](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Manipulation-Augmented) | Training | CC-BY-4.0 | — | 1K Franka Panda demos with Cosmos Transfer1 domain augmentation |
</details>
---
<details>
<summary><strong>Autonomous Vehicles</strong></summary>
*Multi-sensor driving data and synthetic scenarios for training and validating autonomous vehicle systems.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [PhysicalAI-Autonomous-Vehicles](https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles) | Training | NVIDIA AV Dataset License | — | 1,700 hours multi-sensor data from 25 countries, 306K clips |
| [PhysicalAI-Autonomous-Vehicle-Cosmos-Drive-Dreams](https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicle-Cosmos-Drive-Dreams) | SDG | CC-BY-4.0 | Cosmos | 81K synthetic videos with LiDAR and HD-map annotations |
| [PhysicalAI-Autonomous-Vehicle-Cosmos-Synthetic](https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicle-Cosmos-Synthetic) | SDG | CC-BY-4.0 | Cosmos | Cosmos-generated synthetic driving scenarios |
| [PhysicalAI-Autonomous-Vehicles-NuRec](https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NuRec) | Reconstruction | NVIDIA AV Dataset License | — | NuScenes-based reconstruction data |
</details>
---
<details>
<summary><strong>Synthetic Personas / Data Generation</strong></summary>
*Privacy-safe synthetic personas grounded in real-world demographics for sovereign AI development and synthetic data generation pipelines.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-Personas-USA](https://huggingface.co/datasets/nvidia/Nemotron-Personas-USA) | SDG | CC-BY-4.0 | NeMo Data Designer | 1M US personas grounded in Census demographics |
| [Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) | SDG | CC-BY-4.0 | NeMo Data Designer | 1M Japanese personas aligned with regional statistics |
| [Nemotron-Personas-India](https://huggingface.co/datasets/nvidia/Nemotron-Personas-India) | SDG | CC-BY-4.0 | NeMo Data Designer | 3M Indian personas for sovereign AI development |
| [Nemotron-Personas](https://huggingface.co/datasets/nvidia/Nemotron-Personas) | SDG | CC-BY-4.0 | NeMo Data Designer | 100K US personas with 22 fields aligned to Census data |
</details>
---
<details>
<summary><strong>Privacy / PII Detection</strong></summary>
*Synthetic datasets for training named entity recognition models to detect and redact personally identifiable information.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Nemotron-PII](https://huggingface.co/datasets/nvidia/Nemotron-PII) | NER Training | CC-BY-4.0 | GLiNER-PII | 100K synthetic records with 55+ PII/PHI entity types |
</details>
---
<details>
<summary><strong>Safety / Content Moderation</strong></summary>
*Content safety datasets for training guardrail models covering comprehensive risk taxonomies. Powers NemoGuard content safety models.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Aegis-AI-Content-Safety-Dataset-1.0](https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-1.0) | Content Moderation | CC-BY-4.0 | NemoGuard Permissive/Defensive | 11K annotated interactions covering 13 risk categories |
| [Aegis-AI-Content-Safety-Dataset-2.0](https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0) | Content Moderation | CC-BY-4.0 | Llama-3.1-NemoGuard-8B-ContentSafety | Extended safety dataset with 23 violation categories |
| [Nemotron-Content-Safety-Audio-Dataset](https://huggingface.co/datasets/nvidia/Nemotron-Content-Safety-Audio-Dataset) | Audio Safety | CC-BY-4.0 | — | 1.9K audio files from Aegis 2.0 with accent diversity |
</details>
---
<details>
<summary><strong>RAG / Conversational QA</strong></summary>
*Training and evaluation data for retrieval-augmented generation and conversational question answering. Powers ChatQA models.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [ChatRAG-Bench](https://huggingface.co/datasets/nvidia/ChatRAG-Bench) | Evaluation | Other (derived) | — | Benchmark across 10 datasets for document QA and unanswerable detection |
| [ChatQA-Training-Data](https://huggingface.co/datasets/nvidia/ChatQA-Training-Data) | SFT | Other (derived) | ChatQA-1.5 | Training data for ChatQA models from multiple sources |
| [ChatQA2-Long-SFT-data](https://huggingface.co/datasets/nvidia/ChatQA2-Long-SFT-data) | SFT | Other (derived) | ChatQA-2 | 128K long-context training data for ChatQA-2 |
</details>
---
<details>
<summary><strong>Biology / Drug Discovery</strong></summary>
*Protein sequence data for training biological foundation models.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [esm2_uniref_pretraining_data](https://huggingface.co/datasets/nvidia/esm2_uniref_pretraining_data) | Pre-training | CC-BY-4.0 | ESM2-nv | 188M protein sequences for ESM2 |
</details>
---
<details>
<summary><strong>3D / Spatial Intelligence</strong></summary>
*Testing and synthetic data for 3D reconstruction, video generation, and spatial understanding models.*
| Dataset | Usage | License | Model(s) | Description |
|---------|-------|---------|----------|-------------|
| [Lyra-Testing-Example](https://huggingface.co/datasets/nvidia/Lyra-Testing-Example) | Evaluation | CC-BY-4.0 | Lyra | Testing examples for Lyra generative 3D reconstruction |
| [PhysicalAI-SpatialIntelligence-Lyra-SDG](https://huggingface.co/datasets/nvidia/PhysicalAI-SpatialIntelligence-Lyra-SDG) | SDG | CC-BY-4.0 | Lyra | Synthetic data for spatial intelligence models |
| [GEN3C-Testing-Example](https://huggingface.co/datasets/nvidia/GEN3C-Testing-Example) | Evaluation | CC-BY-4.0 | GEN3C | Testing examples for GEN3C video generation |
| [ChronoEdit-Example-Dataset](https://huggingface.co/datasets/nvidia/ChronoEdit-Example-Dataset) | Evaluation | CC-BY-4.0 | ChronoEdit | Temporal reasoning examples for image editing |
</details>
</details>
---
## 💡 Feature Requests & Ideas
Have an idea for improving Nemotron models? Create a [Discussion](https://github.com/NVIDIA-NeMo/Nemotron/discussions) topic for it!
If you have a feature request, feel free to open an [Issue](https://github.com/NVIDIA-NeMo/Nemotron/issues) and tag it as `enhancement`.
Your feedback helps shape the future of Nemotron models!
---
## Documentation
- [Nemotron 3 Super Training Guide](docs/nemotron/super3/README.md) – frontier model training recipe
- [Nemotron 3 Nano Training Guide](docs/nemotron/nano3/README.md) – efficient model training recipe
- [Nemotron Training FAQ](docs/train-models/faq/index.md) – answers to common questions about training and fine-tuning Nemotron models
- [NeMo-Run Configuration](docs/nemo_runspec/nemo-run.md) – execution profiles and job orchestration
- [Data Preparation](docs/nemotron/data-prep.md) – data preparation module
- [Contributing Guidelines](CONTRIBUTING.md) – how to contribute
- [Changelog](CHANGELOG.md) – version history
---
## Contributing
We welcome contributions: examples, recipes, or other tools. Please read the [Contributing Guidelines](CONTRIBUTING.md) before submitting pull requests.
---
## Security
To report any vulnerabilities, please reach out to [security@nvidia.com](mailto:security@nvidia.com)
---
## License
Apache 2.0 License — see [LICENSE](LICENSE) for details.
---
**NVIDIA Nemotron** — Open and efficient models for agentic AI.