Powered byLightning AI
← Managed inference

Nemotron 3 Super 120B

An architecturally different option for agents and long-context work

Start free

Overview

NVIDIA Nemotron 3 Super is a LatentMoE hybrid -- Mamba-2 state-space layers interleaved with MoE and attention blocks, plus a multi-token-prediction head -- with 120B total parameters and 12B active. NVIDIA targets it at agentic workflows, long-context reasoning, tool use and RAG. Documented context runs to 1M tokens; served here at 262K. Released March 2026.

Strengths

  • Not a pure transformer, so its failure modes differ from the Qwen models
  • 262K served context
  • Only 12B active parameters per token

Use cases

  • Agentic workflows and tool use
  • Long-context reasoning and RAG
  • A second opinion when a Qwen model is stuck

Good to know

NVIDIA documents a configurable reasoning mode, but this deployment does not surface thinking content, so treat it as non-thinking. Priced at the top of the chat range, level with qwen3-coder. Covers seven languages against the Qwen models' 201.

API access

One OpenAI-compatible API for every model in the catalog, plus the CLI and the function SDK inside workflows. Sign up free, add a card for $5 in credits, and these requests work:

nemotron-3-super · curl
curl https://gateway.graphn.ai/v1/chat/completions \
  -H "Authorization: Bearer $GRAPHN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nemotron-3-super",
    "messages": [{"role": "user", "content": "Hello"}]
  }'