NVIDIA model product
Nemotron 3 Super is a 120B total / 12B active parameter hybrid Mamba-Attention Mixture-of-Experts model optimized for agentic reasoning, coding, planning, tool calling, and long-context analysis. It introduces LatentMoE (projecting tokens into a compressed latent space for expert routing, enabling 4x more experts at the same inference cost), Multi-Token Prediction for native speculative decoding (up to 3x faster generation), and native NVFP4 pretraining on Blackwell. The hybrid architecture interleaves Mamba-2 layers for linear-time sequence processing with strategically placed Transformer attention layers as global anchors, supporting a 1M-token context window. Pre-trained on 25 trillion tokens and post-trained with multi-environment RL across 21 configurations using NeMo Gym/RL with 1.2 million rollouts. Achieves up to 5x higher throughput than previous Nemotron Super and 2.2x higher throughput than GPT-OSS-120B while maintaining comparable accuracy.
Updated Aug 10, 2026. Default version: Nemotron 3 Super (120B A12B)
Structured fields from the published default version.
This profile uses the latest version under this unique model that has a calculated LLMBoard score. Arena and price are excluded.
Published benchmark records for the scored version Nemotron 3 Super (120B A12B).
| WMT24++ | 0.9 | 1 | 23 | 100.0% | C | |
| RULER | 0.9 | 2 | 4 | 66.7% | C | |
| Bird-SQL (dev) | 0.4 | 5 | 7 | 33.3% | C | |
| Arena-Hard v2 | 0.7 | 6 | 16 | 66.7% | C | |
| LiveCodeBench | 0.8 | 7 | 73 | 91.7% | C | |
| HMMT 2025 | 0.9 | 8 | 33 | 78.1% | C | |
| SciCode | 0.4 | 10 | 18 | 47.1% | C | |
| AA-LCR | 0.6 | 12 | 15 | 21.4% | C | |
| MMLU-ProX | 0.8 | 12 | 32 | 64.5% | C | |
| Multi-Challenge | 0.6 | 12 | 29 | 60.7% | C | |
| IFBench | 0.7 | 13 | 28 | 55.6% | C | |
| Tau2 Airline | 0.6 | 18 | 23 | 22.7% | C | |
| Terminal-Bench | 0.3 | 22 | 25 | 12.5% | C | |
| Tau2 Retail | 0.6 | 24 | 26 | 8.0% | C | |
| MMLU-Pro | 0.8 | 29 | 129 | 78.1% | C | |
| Tau2 Telecom | 0.6 | 29 | 35 | 17.6% | C | |
| SWE-bench Multilingual | 0.5 | 32 | 34 | 6.1% | C | |
| AIME 2025 | 0.9 | 47 | 114 | 59.3% | C | |
| Terminal-Bench 2.0 | 0.3 | 49 | 49 | 0.0% | C | |
| Humanity's Last Exam | 0.2 | 53 | 92 | 42.9% | C | |
| BrowseComp | 0.3 | 54 | 58 | 7.0% | C | |
| GPQA | 0.8 | 70 | 233 | 70.3% | C | |
| SWE-Bench Verified | 0.5 | 86 | 104 | 17.5% | C |
Preference and agent-evaluation signals from published Arena datasets.
The default version has no published Arena rows, or its source alias has not been resolved.
Official vendor API PAYG pricing is summarized first. The table then lists individual provider offerings without treating their minimum as the official price.
| Kenari | nemotron-3-super-120b-a12b | global | N/A | N/A | 262.1K |
Official prices use only the vendor's configured official Provider and positive standard USD PAYG rates. Third-party offers remain explicitly labeled.
All published versions linked to this unique model. The score columns identify the version used by the current overall ranking.
| Nemotron 3 Super (120B A12B) | 59.1 | 120B | 262.1K | 262.1K | Yes | NVIDIA Open Model License Agreement |
A concise description based on the published model registry.
Nemotron 3 Super is a 120B total / 12B active parameter hybrid Mamba-Attention Mixture-of-Experts model optimized for agentic reasoning, coding, planning, tool calling, and long-context analysis. It introduces LatentMoE (projecting tokens into a compressed latent space for expert routing, enabling 4x more experts at the same inference cost), Multi-Token Prediction for native speculative decoding (up to 3x faster generation), and native NVFP4 pretraining on Blackwell. The hybrid architecture interleaves Mamba-2 layers for linear-time sequence processing with strategically placed Transformer attention layers as global anchors, supporting a 1M-token context window. Pre-trained on 25 trillion tokens and post-trained with multi-environment RL across 21 configurations using NeMo Gym/RL with 1.2 million rollouts. Achieves up to 5x higher throughput than previous Nemotron Super and 2.2x higher throughput than GPT-OSS-120B while maintaining comparable accuracy.
Use the benchmark, Arena and pricing sections above as separate evidence. A missing field means the current data snapshot does not support that claim.
Data snapshot: 2026-08-07. Editorial model content is not available in the backend.
Open a comparison with the three ranked models immediately above and below this model.
Recommendations prioritize the same model type and family, then the closest published LLMBoard score.
Common questions about Nemotron 3 Super.
Nemotron 3 Super's default version was released on Mar 11, 2026.
No official standard PAYG price is currently available for Nemotron 3 Super.
Nemotron 3 Super is published under NVIDIA in the model registry.
The default version has a 262.1K token context window.
Yes. The default version is marked as open weight under NVIDIA Open Model License Agreement .
1 published provider offerings are linked to the default version.
Nearby ranked alternatives include LongCat Flash Thinking, Nemotron Nano 9B, Kimi K2 Thinking.