Nemotron 3 Super

Changes from previous version

Scales Nemotron 3 Nano from 30B total and 3B active parameters to 120B total and 12B active parameters, adding Multi-Token Prediction for higher-throughput complex agent workflows.

Release Summary

NVIDIA's open 120B-total, 12B-active hybrid MoE targets high-volume agentic reasoning, coding, planning, RAG, and tool use. Its native 1M-token context keeps long multi-agent workflows in memory while Multi-Token Prediction accelerates generation.

Timeline

Other Nemotron versions