NVIDIA Nemotron provides a developer resource hub for open, high-efficiency AI models, targeting agentic AI workflows. It offers transparent training data, weights, and recipes, enabling efficient inference, fine-tuning, and deployment across diverse compute platforms and frameworks.

How It Works

Nemotron models employ novel architectures like hybrid Mamba-Transformer MoE and Mamba-2. The repository includes usage cookbooks for deployment and model interaction, plus detailed use-case examples for agentic workflows, RAG, and tool integration. Future training recipes will offer end-to-end reproducible pipelines using NeMo ecosystem tools. This approach prioritizes transparency, efficiency, and adaptability across compute tiers and serving frameworks (NeMo, TensorRT-LLM, vLLM, SGLang, NIM).

Quick Start & Requirements

This repo acts as a resource hub, not a standalone package. Integration occurs via frameworks like NeMo, TensorRT-LLM, vLLM, or SGLang. Hardware needs (GPU, CUDA) vary by model; one model requires a single H200 GPU. Key resources: Nemotron Developer Page, Nemotron Research Hub, Nemotron Datasets.

Highlighted Details

• Models feature advanced architectures (e.g., hybrid MoE, Mamba-2), large context windows (up to 1M tokens), and specialized capabilities like controllable reasoning, vision-language (VLM), and document parsing.
• Specific models include Nemotron-3-Nano (3.6B active MoE, 1M context), Llama-3.3-Nemotron-Super-49B (128K context, NAS-optimized), Nemotron-Nano-12B-v2-VL (VLM), and Llama-3.1-Nemotron-Safety-Guard-8B (multilingual moderation).
• Use-case examples cover agentic workflows, RAG, tool integration, and production patterns.

Maintenance & Community

Contributions are welcomed via pull requests following Contributing Guidelines. Community feedback and feature requests can be submitted and voted on via the Nemotron Ideas Portal.

Licensing & Compatibility

Licensed under the permissive Apache 2.0 License, allowing broad compatibility, including commercial use.

Limitations & Caveats

Full training pipelines ("Training Recipes") are marked "Coming Soon." Documentation links for "Nemotron-Parse" are also pending.