Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

TL;DR

Nemotron 3 Super is a large hybrid Mixture-of-Experts model with innovative architecture and training techniques, achieving high efficiency and accuracy, and is open-sourced for research use.

Contribution

Introduces LatentMoE architecture, NVFP4 pre-training, and MTP layers for acceleration, setting new standards in model efficiency and open access.

Findings

Supports up to 1 million tokens context length.

Achieves 2.2x and 7.5x higher inference throughput than comparable models.

Pre-trained on 25 trillion tokens with open-sourced datasets and checkpoints.

Abstract

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.