Nemotron 3 Super · NVIDIA · Model History

Claims and evidence

Announced 15 December 2025Announced with the Nemotron 3 family, then described as about 100B parameters with up to 10B active and expected in the first half of 2026. Primary source[3]
Released 11 March 2026Other sources give: 10 March 2026 [4]Published Publication date shown on the NVIDIA Nemotron research page; the launch post and the model card give March 11, 2026. Primary source[1] [2]Model Summary: Release Date; Release Date

Status AvailableNVIDIA has no deprecation page for its open models; the weights were still downloadable from the NVIDIA organisation on Hugging Face on 2026-10-01. Primary source[2]

Successor of Llama-3.3-Nemotron-Super-49B-v1.5 EditorialEditorial link along the Super tier: the launch post compares the model with the previous Nemotron Super model without naming a version; Llama-3.3-Nemotron-Super-49B-v1.5 was the most recent Super release before it. Primary source[1]Hybrid Architecture

Change · Architecture Uses a hybrid Mamba-Transformer latent mixture-of-experts design with multi-token prediction, activating 12B of its 120B parameters per token.Compared with Llama-3.3-Nemotron-Super-49B-v1.5 Primary source[1]Hybrid Architecture
Change · Context length Supports a context window of up to 1M tokens.Compared with Llama-3.3-Nemotron-Super-49B-v1.5 Primary source[2]Model Summary: Context Length
Change · Efficiency NVIDIA reports higher throughput than the previous Nemotron Super model and says the model runs in NVFP4 precision on Blackwell GPUs.Compared with Llama-3.3-Nemotron-Super-49B-v1.5 Primary source[1]
Change · Training data NVIDIA published over 10 trillion tokens of pre- and post-training data and 15 reinforcement learning environments with the model.Compared with Llama-3.3-Nemotron-Super-49B-v1.5 Primary source[1]Open Weights, Data and Recipes

Input text Primary source[2]Input
Output text Primary source[2]Output
Feature Reasoning modeReasoning can be switched on or off through the chat template. Primary source[2]Model Summary: Reasoning Mode
Feature Tool use Primary source[1]Use in Agentic Systems [2]Model Summary: Best For
Feature MultilingualEnglish, French, German, Italian, Japanese, Spanish and Chinese. Primary source[2]Model Summary: Supported Languages
Open weights YesThe launch post calls the licence permissive; the model card names the NVIDIA Nemotron Open Model License. Primary source[1]Open Weights, Data and Recipes [2]License/Terms of Use
Context window 1M tokensStated as up to 1M tokens. Primary source[2]Model Summary: Context Length [1]
Parameters 120B (12B active) Primary source[2]Model Summary: Total Parameters [1]
Access Open-weights download, API, cloud partnerAt launch: Hugging Face, build.nvidia.com, Perplexity and OpenRouter, Google Cloud Vertex AI and Oracle Cloud Infrastructure, and an NVIDIA NIM microservice; Amazon Bedrock and Microsoft Azure were announced as coming soon. Primary source[1]Availability

No known successor.

Not derived from another model.

None recorded.

None recorded.

None.

Same dates as Nemotron 3 Super.

Variant Inline variant in this record.Post-trained model, published in BF16 and as FP8 and NVFP4 quantized checkpoints. Primary source[4]Open Source: Checkpoints [2]

Also known as: NVIDIA-Nemotron-3-Super-120B-A12B-BF16, NVIDIA-Nemotron-3-Super-120B-A12B-FP8, NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

Same dates as Nemotron 3 Super.

Variant Inline variant in this record.Pre-trained base checkpoint. Primary source[4]Open Source: Checkpoints [2]Training Methodology: Stage 1

Also known as: NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16

AI Radar coverage starts in June 2026; no coverage linked yet.