---
title: "Nemotron 3 Super"
type: model
id: "nemotron-3-super"
provider: "NVIDIA"
model_type: "open-source"
api_model_id: "nvidia/nemotron-3-super-120b-a12b"
release_date: "2026-03"
description: "NVIDIA's open-weight MoE model with 120B total parameters and 12B active. Designed to fit on hardware most companies already own. Serves as the anchor for NVIDIA's agent toolkit strategy, optimized for TensorRT-LLM and NIM."
last_updated: "2026-04-30"
last_verified: "2026-04-30"
knowledge_cutoff: "2026-02"
availability_status: "available"
deprecated: false
tool_schema_format: "openai-compatible"
pricing_confidence: "high"
model_listing_confidence: "high"
benchmark_confidence: "high"
context_window: "1M tokens"
website: "https://build.nvidia.com"
license: "NVIDIA Open Model License"
modality:
- "text"
tags:
- "nvidia"
- "open-source"
- "text"
pricing:
  input: "Free (open weights)"
  output: "Free (open weights)"
  free: true
  note: "Also via NVIDIA NIM API"
benchmarks:
  reasoning: 80
  coding: 82
  math: 78
  writing: 79
  multilingual: 78
  speed: 88
capabilities:
- "function_calling"
- "structured_output"
- "streaming"
- "tool_search"
- "long_context"
- "reasoning"
sources:
- title: "NVIDIA Nemotron 3 Super model card"
  url: "https://build.nvidia.com/nvidia/nemotron-3-super-120b-a12b/modelcard"
- title: "NVIDIA Nemotron 3 Super Hugging Face model card"
  url: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
benchmark_sources:
- title: "NVIDIA Nemotron 3 Super model card"
  url: "https://build.nvidia.com/nvidia/nemotron-3-super-120b-a12b/modelcard"
parameters: "120B total (12B active)"
hardware_requirements: "2x H100 80GB (FP8); 8x H100 80GB (BF16)"
best_for:
- "Enterprise deployment"
- "NVIDIA hardware optimization"
- "Agent workflows"
- "Production inference"
---

# Nemotron 3 Super

NVIDIA's play to own the enterprise AI stack from GPU to model. Nemotron 3 Super is a 120B MoE model with 12B active parameters, purpose-built for TensorRT-LLM and NIM -- NVIDIA's inference and deployment frameworks. If your company already runs NVIDIA hardware (and statistically, it does), this model is optimized for it in ways that generic open models are not.

Coding at 82 and reasoning at 80 anchor the benchmark profile, with math (78), writing (79), and multilingual (78) rounding out a solidly mid-tier picture. Speed at 88/100 reflects the TensorRT optimization. The official deployment targets are heavier than the original draft implied: plan for multi-H100 infrastructure rather than a single workstation GPU.

The agent toolkit strategy is the bigger story. NVIDIA is positioning Nemotron 3 Super as the default model for their agent workflow ecosystem, meaning it gets first-class support for tool calling, multi-step planning, and agentic deployment patterns. If you are building AI agents on NVIDIA infrastructure, the integration is seamless.

The NVIDIA Open Model License is not Apache 2.0 -- it is more restrictive on redistribution and modification. Community fine-tuning is still thin compared to Llama or Qwen, and the model clearly performs best on NVIDIA hardware, limiting portability.

**When to pick something else:** NVIDIA's own Nemotron-Cascade 2 beats it on coding (90 vs 82) and math (92 vs 78) while being dramatically smaller. For vendor-neutral deployment, Qwen 3.5 or Mistral Small 3 under Apache 2.0 avoid NVIDIA lock-in. Nemotron 3 Super is for NVIDIA-committed enterprises building agent infrastructure.