Overview
Run NVIDIA Nemotron™ 3 Super 120B A12B with high-performance inference. Quickly deploy NVIDIA’s hybrid Mamba-Transformer MoE with 12B active parameters for agentic reasoning.
Explore additional models
Frequently
asked questions
Nemotron 3 Super 120B A12B is good for enterprise agents, coding assistance, and multi-step reasoning that exceeds what a 3B-active model handles well but doesn’t justify a frontier-scale model. It is a 123.6B hybrid Mamba-Transformer mixture-of-experts with about 12B active parameters, post-trained by NVIDIA for agentic planning, tool calling, and multi-turn task recovery.
Nemotron 3 Super 120B A12B is a text-only reasoning model, while Nemotron 3.5 Super VL is the newer vision-language model in the same size class, adding image input for document, chart, and visual understanding. Both activate about 12B parameters per token and are available for Serverless Inference in Crusoe Intelligence Foundry. Choose Super 120B A12B for text-based agents and reasoning, and Super VL when inputs include images.
“Super” is NVIDIA’s name for the mid-size tier of the Nemotron model family, sitting above the efficiency-focused Nano models. Super models trade some speed for more reasoning depth, with more active parameters per token, which suits enterprise agents and multi-step tasks that a Nano model can’t handle well. On Crusoe, the Super tier includes Nemotron 3 Super 120B A12B and Nemotron 3.5 Super VL.
Nemotron 3 Super 120B A12B is the strongest Nemotron model for text reasoning on Crusoe, with four times the active capacity of Nemotron 3 Nano 30B A3B for agents that need more depth. Its hybrid Mamba-Transformer design keeps long-context inference efficient across the full 262,144-token window, so long tool traces and document sets stay in a single request.
Nemotron 3 Super 120B A12B is available on Crusoe Cloud, no infrastructure set up required. It is currently available for high-performance Serverless Inference, but not Self-Serve Deployments or Serverless Fine-Tuning. For a Nemotron model you can fine-tune and deploy on a dedicated GPU, use Nemotron 3.5 Lightning 30B A3B. For an SLA-backed endpoint, contact us about Tailored Deployments.
The “A12B” in Nemotron 3 Super 120B A12B stands for about 12 billion active parameters. The model is a mixture-of-experts with roughly 120B total parameters, but it routes each token through only about 12B of them. That gives it 120B-class knowledge while computing like a much smaller model, keeping latency and serving costs well below a dense model of the same size.


%204.png)
