Nemotron 3 Super 120B A12B (Reasoning)

Nemotron 3 Super is a 120B total / 12B active parameter hybrid Mamba-Attention Mixture-of-Experts model optimized for agentic reasoning, coding, planning, tool calling, and long-context analysis. It introduces LatentMoE (projecting tokens into a compressed latent space for expert routing, enabling 4x more experts at the same inference cost), Multi-Token Prediction for native speculative decoding (up to 3x faster generation), and native NVFP4 pretraining on Blackwell. The hybrid architecture interleaves Mamba-2 layers for linear-time sequence processing with strategically placed Transformer attention layers as global anchors, supporting a 1M-token context window. Pre-trained on 25 trillion tokens and post-trained with multi-environment RL across 21 configurations using NeMo Gym/RL with 1.2 million rollouts. Achieves up to 5x higher throughput than previous Nemotron Super and 2.2x higher throughput than GPT-OSS-120B while maintaining comparable accuracy.