Skip to content
NEWLive updates feed from Hugging Face, arXiv and GitHub  See what changed →

nvidia · nemotron_h

NVIDIA-Nemotron-3-Super-120B-A12B-BF16

NVIDIA-Nemotron-3-Super-120B-A12B-BF16 by nvidia: a 123.6B-parameter mixture-of-experts open-weight model under the other license.

otherMixture of experts262K contextAuto-imported · not yet reviewed

1,275,676 downloads · 426 likes · Model card · synced 2026-09-19

Specifications

Parameters
123.6B (5.3B active)
Architecture
moe · nemotron_h
Layers
88
Hidden size
4,096
KV heads · head dim
2 · 128
Vocabulary
131,072
Native dtype
Context window
262,144 tokens
Released
2026-03-10
Last modified on Hub
2026-08-25

Features & licensing

License
other
Commercial use
No
Modalities
text
Tool / function calling
No
Reasoning mode
No
Languages
EN, FR, ES, IT, DE, JA, ZH
Quantised variants on Hub
GGUF, bitsandbytes
Pipeline
text-generation
transformerssafetensorsnemotron_htext-generationnvidiapytorchnemotron-3latent-moemtpconversationalcustom_codeenfresitdejazheval-resultsendpoints_compatible

Task fit

Editorial scores (0–100) used by the recommendation engine.

  • Chat70
  • RAG / Q&A70
  • Code70
  • Summarisation70
  • Extraction70
  • Agents70

Hardware to run NVIDIA-Nemotron-3-Super-120B-A12B-BF16

Weights need about 272 GB at FP16, 143 GB at INT8 and 82 GB at INT4. The KV cache adds roughly 90 MB per 1,000 tokens per request. For a reference workload of 5 requests per second with 1,500 input and 300 output tokens, total GPU memory is around 110 GB, which fits on 2 × NVIDIA A100 (80 GB).

GPU configurations at the reference workload
ConfigurationVRAMUtilisationEst. first tokenCloud / month
5 × NVIDIA L4 (24 GB)120 GB92 %~0.5 s$2,920
3 × NVIDIA L40S (48 GB)144 GB76 %~0.2 s$4,161
2 × NVIDIA A100 (80 GB)recommended160 GB69 %~0.1 s$4,672
2 × NVIDIA H100 (80 GB)160 GB69 %~0.0 s$6,570

Estimates only. Run an assessment for your own traffic.

Auto-imported from Hugging Face. Architecture, license and feature data are synced from the model repository; the task-fit scores below are provisional defaults until a curator reviews them.