NVIDIA Nemotron 3: Efficient and Open Intelligence

2025-12-23

AI Summary

This paper addresses the joint challenges of enhancing agent capabilities, long-context reasoning, and throughput optimization in large language models. To this end, we propose OpenAgent—a family of open-source large models designed for efficiency, openness, and multi-scenario adaptability. Our method introduces a novel Mamba-Transformer hybrid Mixture-of-Experts (MoE) architecture: (i) LatentMoE, a first-of-its-kind mechanism to improve expert quality via latent routing; (ii) MTP (Multi-Token Prediction) layers to accelerate autoregressive generation; and (iii) fine-grained inference budget control. The series comprises three variants—Nano (optimal accuracy-cost trade-off), Super (high-concurrency collaborative agents), and Ultra (state-of-the-art inference performance)—all supporting unified 1M-token context windows and multi-environment reinforcement fine-tuning. We fully open-source model weights, training code, and datasets, and integrate NVFP4 quantization with end-to-end training pipelines.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Multiagent Systems: Modeling other Agents

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Problem

• Introduces Nemotron 3 models for strong agentic and reasoning capabilities
• Uses hybrid architecture for high throughput and long context handling
• Applies novel training for improved model quality and efficient inference

Innovation

• Mixture-of-Experts hybrid Mamba-Transformer architecture
• LatentMoE and NVFP4 training for improved quality
• Multi-environment reinforcement learning post-training for reasoning