Nemotron 3 Super
Released 2026-03 · reasoning · 1.0M tokens · 9 benchmarks · Open weight
Editorial notes
Lanzado 11 marzo 2026 en GTC 2026. Hybrid Mamba-Transformer MoE 120B / 12B activos. LatentMoE architecture. 5x throughput NVFP4 Blackwell. Supera GPT-OSS-120B con +10% throughput/GPU. #1 DeepResearch Bench.
Spec sheet
- Company
- Nvidia
- Country
- US
- Type
- reasoning
- Release
- 2026-03
- Context
- 1.0M tokens
- Params total
- 120B
- Params active (MoE)
- 12B
- License
- nvidia-open-model
- Quants
- BF16, Q8_0, Q5_K_M, Q4_K_M
- Pricing (openrouter)
- $0.09/$0.45/M
- Slug
- nemotron-3-super
Quick install
2 toolsollama run nemotron-3:superNote: ~70 GB VRAM Q4_K_M
vllm serve nvidia/Nemotron-3-Super --tensor-parallel-size 2Benchmarks (9)
Reasoning 4
- 86.0MMLUMassive Multitask Language Understanding — 57 academic subjects, ~16K questions.
- 83.3MMLU-ProMMLU upgraded with harder questions and 10 answer options.
- 79.4GPQA-DiamondGraduate-level Physics, Chemistry, Biology — PhD-level questions.
- 17.4Humanitys-Last-ExamThe hardest known benchmark — novel academic problems.
Coding 3
Cite this model
BibTeX · APA
BibTeX
@misc{frontier-nemotron-3-super,
title = {Nemotron 3 Super},
author = {{Nvidia}},
year = {2026},
note = {Frontier Benchmarks AI atlas. Accessed 2026-05-08},
url = {https://frontierbenchmarks.com/models/nemotron-3-super}
} APA
Nvidia (2026). Nemotron 3 Super [Large language model]. Frontier Benchmarks AI. Retrieved 2026-05-08, from https://frontierbenchmarks.com/models/nemotron-3-super
Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.