News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnNemotron 3 Super

Nemotron 3 Super

NVIDIA's 120B hybrid MoE language model with 12B active parameters and 1M-token context

Nemotron 3 Super is a large language model released by NVIDIA on 11 March 2026, with 120 billion total parameters of which 12 billion are active per token. It uses a hybrid LatentMoE design and targets agentic work, long-context reasoning and high-volume workloads such as IT ticket automation.

Share on WhatsApp

  • 120B total parameters, 12B active per token
  • Context window of up to 1M tokens
  • Hybrid LatentMoE: Mamba-2, MoE and attention layers
  • Pre-trained in NVFP4, the first Nemotron 3 model at that precision
  • Released 11 March 2026 under the NVIDIA Nemotron Open Model License

What it is

Nemotron 3 Super is a general-purpose language model from NVIDIA, put out on 11 March 2026 as part of the open Nemotron family. It carries 120 billion parameters in total but draws on only 12 billion of them for each token, which keeps running costs down. The company positions it for collaborative agents, long-context reasoning and heavy workloads such as IT ticket automation, as well as chatbots, retrieval-augmented generation and complex instruction following. It handles English, French, German, Italian, Japanese, Spanish and Chinese, and a flag in the chat template switches its reasoning trace on or off. NVIDIA says the model is ready for commercial use.

How it came about

NVIDIA dates the model's development to December 2025 through March 2026, with pre-training data cut off in June 2025 and post-training data in February 2026. The weights appeared on Hugging Face on 11 March 2026, alongside a base checkpoint and a separate checkpoint carrying an updated MTPv2 speculative decoding head. A technical report for the model is published by NVIDIA's research labs. The company describes Nemotron as a line of open models that ship with weights, training data and recipes.

Inside the model

Tokens are mapped to a smaller latent space before expert routing, which NVIDIA says raises accuracy per byte. Super was pre-trained in NVFP4, a 4-bit number format, the first Nemotron 3 model trained this way; most linear layers use NVFP4 for weights, activations and gradients, while latent projections, prediction heads, attention projections and embeddings stay in BF16 or MXFP8 for stability. Multi-Token Prediction heads share weights, improving the training signal and enabling speculative decoding.

What it can do

NVIDIA reports scores across knowledge, maths, code and long-context tests. MMLU-Pro is 83.73, AIME25 without tools 90.21, GPQA 79.23 without tools and 82.70 with tools, and LiveCodeBench v5 81.19. On SWE-Bench via OpenHands it scores 60.47. Long-context retention is reported as 96.30 on RULER at 256k tokens, 95.67 at 512k and 91.75 at 1M. The company's table also sets it beside Qwen3.5-122B-A10B and GPT-OSS-120B, where it leads on some tests and trails on others, for example MMLU-Pro and GPQA. Most results came from NVIDIA's own Nemo evaluation tools.

How to use it

The BF16 weights are downloadable from Hugging Face under the NVIDIA Nemotron Open Model License, and NVIDIA states the model is ready for commercial use. Running the NIM container brings in the NVIDIA Software License Agreement and product-specific terms on top. Serving needs at least eight H100 80GB GPUs; an NVFP4 build is offered for a single B200 or DGX Spark. NVIDIA advises temperature 1.0 and top_p 0.95 for every task and backend.

Questions people ask about Nemotron 3 Super

What is LatentMoE?

It is NVIDIA's name for the model's design, in which tokens are projected into a smaller latent space before being routed to experts. Mamba-2 layers are interleaved with mixture-of-experts layers and some attention layers. NVIDIA says this improves accuracy per byte.

How much hardware do I need to run it?

The BF16 checkpoint needs a minimum of eight H100 GPUs with 80GB each. NVIDIA also publishes an NVFP4 version that runs on a single B200 or a DGX Spark system.

Can I use it commercially?

Yes. NVIDIA states the model is ready for commercial use, under the NVIDIA Nemotron Open Model License. If you deploy through the NIM container, the NVIDIA Software License Agreement and product-specific terms for AI products also apply.

Which languages does it support?

English, French, German, Italian, Japanese, Spanish and Chinese. NVIDIA describes it as a general-purpose reasoning and chat model intended for English, code and supported multilingual contexts.

Does it always show its reasoning?

No. The model normally writes a reasoning trace before its final answer, but a flag in the chat template, enable_thinking, turns that behaviour on or off.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Nemotron 3 Super yet, so the facts on this page come from the sources listed above: huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.