What is Nemotron 3 Super?
Nemotron 3 Super is the middle model of NVIDIA's Nemotron 3 family, between Nano and Ultra, and it is still current. It mixes Mamba-2, attention and mixture-of-experts layers, uses 12B of its 120B parameters per token, and reads up to 1M tokens of text, with reasoning you can switch off. Its pre-training data stops at June 2025.
NVIDIA built it for the harder steps in agent systems and releases the weights, over 10 trillion tokens of training data and the recipes. Its 4-bit build runs on a single B200 or DGX Spark. It reads no images.
Where does Nemotron 3 Super fit in the Nemotron family?
It is the middle tier: bigger and more accurate than Nano, far lighter than Ultra, and NVIDIA's pick for multi-agent work on a single data center GPU.
- Nemotron 3 Nano
- A 30B model with 3B active, for cheap, specialised sub-agents; Nano Omni adds images, audio and video.
- Nemotron 3 Super
- 120B with 12B active, for reasoning and tool calling in multi-agent apps.
- Nemotron 3 Ultra
- 550B with 55B active, for the hardest reasoning, on at least 4 B200 or 8 H100 GPUs.
- Nemotron 3.5 Lightning
- A newer 30B model with 3B active, built for fast, repetitive steps in always-on agents.
Who is Nemotron 3 Super best for?
Nemotron 3 Super is best for multi-agent systems that need long context on their own GPUs, teams that want to fine-tune or audit a fully open model, and owners of NVIDIA Blackwell hardware.
Multi-agent systems
NVIDIA built it for planning and tool calls in multi-agent apps, with its small Nano model handling the simple steps.
Long documents and codebases
A 1M-token window lets an agent keep a whole codebase or a stack of reports in memory, and it scores 96.09 on RULER at 512K tokens in NVIDIA's table.
Teams that fine-tune and audit
The weights, training data and recipes are published under a licence that allows commercial use and derivative models.
NVIDIA Blackwell owners
It was trained in NVFP4, so the 4-bit build runs on one B200 or DGX Spark, and NVIDIA says it runs up to 4x faster on Blackwell than FP8 on Hopper.
What's included with Nemotron 3 Super?
Switchable reasoning, built-in speculative decoding, three checkpoint sizes, tool calling with structured output, open training data and support for the standard inference servers.
- Reasoning you can switch
- Turn thinking on or off per request, use a low-effort mode, or cap the reasoning with a token budget.
- Multi-token prediction
- Built-in speculative decoding predicts several tokens at once; NVIDIA puts the gain at 3x faster inference.
- Three checkpoints
- BF16, FP8 and NVFP4 builds on Hugging Face, from eight H100 GPUs down to one B200 or DGX Spark.
- Tool calling and structured output
- Function calling and structured output are supported, and NVIDIA ships setup guides for OpenCode, OpenHands and OpenClaw.
- Open training data
- Over 10 trillion tokens of datasets, 15 reinforcement learning environments and the evaluation recipes are public.
- Standard servers
- Runs on vLLM, SGLang and TensorRT-LLM with an OpenAI-compatible API, or as an NVIDIA NIM container.
How much does Nemotron 3 Super cost?
Nemotron 3 Super is free: the weights download at no cost with commercial use allowed and NVIDIA runs a free trial API, while NVIDIA's supported NIM container for production needs an AI Enterprise licence at $4,500 per GPU per year.
- NVIDIA Nemotron Open Model License, commercial use allowed
- BF16, FP8 and NVFP4 checkpoints on Hugging Face
- You pay only for the GPUs to run it
- For prototyping, under NVIDIA's API trial terms
- Prompts and replies are recorded to improve NVIDIA products
- Needs a free NVIDIA API key
- Needed only to run NVIDIA's NIM container in production
- 90-day free trial licence
- Education, Inception and Connect members pay $1,125
NVIDIA sells no per-token plan for Nemotron 3 Super. Partner hosts listed on NVIDIA's model page set their own serverless prices: OpenRouter $0.08 / $0.45 and DigitalOcean $0.30 / $0.65 per million input / output tokens. NVIDIA AI Enterprise also comes as a perpetual licence with five years of support at $22,500 per GPU, or at $1 per GPU per hour plus instance costs on cloud marketplaces. Full pricing on build.nvidia.com. We checked these prices on .
How do readers rate Nemotron 3 Super?
No one has rated Nemotron 3 Super yet. Be the first.
Sign in to rate. One rating per account.
Nemotron 3 Super pros and cons
It is fully open, holds its speed on long prompts and runs on one B200, but it is text only and many users find Qwen and Gemma models smarter.
Pros
- Fully open: NVIDIA publishes the weights, over 10 trillion tokens of training data and the recipes, under a licence that allows commercial use.
- Local users say it keeps its speed and focus deep into long prompts; one measured about 30 tokens a second past 100K tokens, while Qwen3.5-122B fell to about 10 by 50K.
- One server is enough: the NVFP4 build runs on one B200 or DGX Spark, and the FP8 build on two H100 GPUs.
- Some teams pick it as a US-made open model that risk-averse customers will sign off on.
Cons
- Text only: it reads no images, which users weighing it against Qwen3.5-122B called the main trade-off.
- Many local users rate Qwen3.6 or Gemma 4 as smarter, and some saw no speed gain over Qwen3.5-122B on their own GPUs.
- Testers report slips in agent work, from failed basic workflow steps to syntax errors and a system prompt leaked into replies.
- It refuses some borderline creative requests that other open models answer.
What are the best Nemotron 3 Super alternatives?
Three tools on Recaply do Nemotron 3 Super's job. All three are open source, so you can run them on your own server.
NVIDIA Nemotron 3 UltraOpen SourceChoose it when accuracy on the hardest planning steps matters more than hardware; it has 55B active parameters out of 550B and needs at least 4 B200 or 8 H100 GPUs.
Mistral Small 4Open SourceChoose it when you need image input and an Apache 2.0 licence at the same size; it is a 119B mixture of experts with 6B active and a 256K context, at $0.15 / $0.6 per million tokens on Mistral's API.
Qwen3.6-27BOpen SourceChoose it when you want a smaller Apache 2.0 model that reads images and video; several local users rated it smarter in their own tests, and Alibaba's API serves it at $0.60 / $3.60 per million tokens.
Questions about Nemotron 3 Super
What is Nemotron 3 Super?
Nemotron 3 Super is NVIDIA's free, open 120B mixture-of-experts model with 12B active parameters and a 1M-token context, built to plan and call tools inside multi-agent systems.
Who is Nemotron 3 Super for?
Nemotron 3 Super is best for multi-agent systems that need long context on their own GPUs, teams that want to fine-tune or audit a fully open model, and owners of NVIDIA Blackwell hardware.
Is NVIDIA's free Nemotron 3 Super API meant for production?
No. The free endpoint on build.nvidia.com is for prototyping under NVIDIA's API trial terms, and it records prompts and replies. For production, host the open weights yourself, use a partner host such as OpenRouter, or license NVIDIA's NIM container.
How much does Nemotron 3 Super cost?
Nemotron 3 Super's paid plans are NVIDIA AI Enterprise at $4,500 per GPU per year.
Is there an open source alternative to Nemotron 3 Super?
Yes. NVIDIA Nemotron 3 Ultra, Mistral Small 4 and Qwen3.6-27B are open source tools that do the same job, so you can run your own.
What are the best Nemotron 3 Super alternatives?
The closest alternatives on Recaply are NVIDIA Nemotron 3 Ultra, Mistral Small 4 and Qwen3.6-27B.
Where can I try Nemotron 3 Super?
On its official website, blogs.nvidia.com.
Can I use Nemotron 3 Super commercially?
Yes. NVIDIA releases it under the NVIDIA Nemotron Open Model License, which allows commercial use and derivative models and claims no ownership of what the model writes. If you redistribute it, you must pass on the licence and keep NVIDIA's notices.
What hardware does Nemotron 3 Super need?
The 4-bit NVFP4 build runs on one B200 or one DGX Spark. The FP8 build needs two H100 80GB GPUs or one B200, and the full BF16 build needs eight H100 80GB GPUs or two B200s.
Can Nemotron 3 Super read images?
No. Nemotron 3 Super takes text in and writes text out. For images, audio and video, NVIDIA offers Nemotron 3 Nano Omni, a smaller multimodal model in the same family.
Our verdict
Nemotron 3 Super is a sound pick for teams that want a fully open model with a long, usable context on their own NVIDIA GPUs.
NVIDIA publishes the weights, data and recipes under a licence that allows commercial use, and users confirm it keeps its speed past 100K tokens.
- Skip it for
- Work that needs image input, or the smartest open model at this size, where Qwen and Gemma models often test better.
- Check first
- Run your own agent tasks against Qwen3.6-27B or Mistral Small 4 before you commit GPUs.
Nemotron 3
Nemotron 3 Nano Omni
Nemotron 3.5 Lightning
Ring-2.6-1T
SubQ
Mistral Medium 3.5