> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/nemotron.md). # Nemotron #### NVIDIA Nemotron: Open Models for Agentic AI **What is Nemotron?** NVIDIA Nemotron is a family of highly efficient, open, multimodal models, datasets, and technologies built for long-running, specialized agentic AI systems. Designed for advanced reasoning, coding, visual understanding, safety, speech, and information retrieval, Nemotron models help agents complete complex tasks faster with leading accuracy at low cost. Training data, model weights, and technical reports for recreating these models are all openly available on Hugging Face. *** **Architecture** The Nemotron 3 family uses a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture that provides best-in-class throughput while achieving better or on-par accuracy compared to standard Transformers. All Nemotron 3 models support context lengths up to 1 million tokens and are post-trained using multi-environment reinforcement learning — enabling reasoning, multi-step tool use, and granular reasoning budget control at inference time. *** **Model Tiers** **Nano** — Nemotron 3 Nano is a 3.2B active parameter (31.6B total) model. It delivers 4x higher throughput than Nemotron 2 Nano and is optimized for cost-efficient, high-accuracy specialized sub-agents. On a single H200 with 8K input and 16K output settings, it provides 3.3x higher throughput than Qwen3-30B-A3B and 2.2x higher than GPT-OSS-20B. **Super** — Nemotron 3 Super is a 120B total parameter model with 12B active parameters, designed for complex agentic AI systems at scale. It delivers up to 5x higher throughput and up to 2x higher accuracy than the previous Nemotron Super generation. It features a 1-million-token context window to retain full workflow state and prevent goal drift. On NVIDIA Blackwell in NVFP4 precision, it runs up to 4x faster than FP8 on Hopper with no accuracy loss. It holds the top position on the Artificial Analysis leaderboard for efficiency and openness among models of its size, and powered NVIDIA's AI-Q research agent to the number one position on DeepResearch Bench and DeepResearch Bench II. **Ultra** — The largest model in the family, designed for applications demanding the highest reasoning accuracy for complex, long-horizon agentic tasks. *** **Nemotron 3 Nano Omni — Multimodal** Nano Omni is an open multimodal model unifying video, audio, image, and text understanding in a single reasoning loop. With a 30B-A3B hybrid MoE architecture incorporating vision and audio encoders, it eliminates the need for separate perception models. It achieves 9x higher throughput than other open omni models with the same interactivity, topping six leaderboards for complex document intelligence and video and audio understanding. Key agentic use cases include computer use agents for navigating graphical user interfaces, document intelligence for enterprise analysis and compliance, and audio-video reasoning for customer service and monitoring workflows. Companies including Foxconn, Palantir, and H Company have adopted it, with Dell Technologies, DocuSign, Oracle, and others evaluating it. *** **Open Data Stack** With Nemotron 3 Nano and Super, NVIDIA released what it describes as the most comprehensive open data stack for text-based agentic AI: over 10 trillion pretraining tokens, over 40 million post-training samples, more than 20 reinforcement learning environment configurations, and full training recipes — all openly available. *** **Specialized Models in the Ecosystem** Beyond the core language models, the Nemotron ecosystem includes speech models (Parakeet family for ASR with state-of-the-art word error rates across 25 languages), vision-language models (Nemotron Nano VL 12B and Llama Nemotron Nano VL 8B for document intelligence and OCR), and a full retrieval-augmented generation stack covering extraction, embedding, and reranking. *** **Deployment** Models are deployable via vLLM, SGLang, Ollama, and llama.cpp on any NVIDIA GPU. They are available as NVIDIA NIM microservices, through Hugging Face, build.nvidia.com, Perplexity, and OpenRouter. Cloud providers include Google Cloud Vertex AI, Oracle Cloud Infrastructure, Amazon Web Services, and Microsoft Azure. On-premises deployment is supported through Dell Technologies and HPE. *** **Strengths** * Hybrid Mamba-Transformer MoE architecture for best-in-class throughput * 1M token context window for long agentic workflows * Full openness: weights, datasets, training recipes, and RL environments all public * Comprehensive multimodal coverage: language, vision, audio, video, speech, retrieval * Optimized for NVIDIA Blackwell with NVFP4 precision for maximum efficiency * Broad cloud and on-premises deployment partner network **Limitations** * Deeply tied to NVIDIA hardware for best performance * Most efficient on Blackwell; older GPU architectures see reduced gains * Primarily English-focused in language coverage * Agentic multi-model system complexity may increase deployment overhead