nemotron-3-nano:cloud

Cloud tags

Key features

• Hybrid architecture — 23 Mamba-2 and MoE layers plus 6 Attention layers; each MoE layer includes 128 experts plus 1 shared expert, with 6 experts activated per token.
• Unified reasoning/non-reasoning — first generates a reasoning trace, then concludes with a final response; configurable via a chat-template flag.
• Lightweight — 3.5B active parameters out of 30B total keeps it efficient for on-device-class workloads.
• Multilingual — English, German, Spanish, French, Italian, and Japanese.

Best practices

• Reasoning traces — allow reasoning traces for higher-quality solutions on hard tasks; disable for faster responses on simple ones.
• Efficiency — the Mamba-2/MoE hybrid keeps memory and compute low relative to dense models.
• License — NVIDIA Open Model License.