NVIDIA Nemotron models, line by line

NVIDIA's Nemotron LLMs: Nemotron 3 Nano, Super, Ultra and 3.5 Lightning hybrid Mamba MoE models, plus Nemotron-H, Nano 2, Cascade, speech and parse.

What the line is

Nemotron is NVIDIA's family of open models, published under the nvidia organization on Hugging Face. The language models share one architectural idea: a hybrid that interleaves Mamba-2 state-space layers with a small number of attention layers, and since Nemotron 3, mixture-of-experts blocks. The cards for the current models quote context lengths up to 1M tokens.

Licenses vary more than in most families, so read the license before you choose. This page lists which one each card names.

Generations in order

Nemotron has one numbered main generation so far, Nemotron 3, with a half-step (3.5) added in 2026. Its models, in release order:

  • Nemotron 3 Nano 30B-A3B (repository created December 4, 2025): a hybrid MoE with 23 Mamba-2 and MoE layers plus 6 attention layers, 128 experts plus one shared expert, 6 experts active per token, 30B total and about 3.5B active parameters. The config sets 262,144 positions. License: NVIDIA Nemotron Open Model License.
  • Nemotron 3 Super 120B-A12B (March 11, 2026): a LatentMoE hybrid with multi-token prediction layers, 120B total and 12B active parameters, with a context of up to 1M tokens per its card. License: NVIDIA Nemotron Open Model License.
  • Nemotron 3 Nano 4B (March 16, 2026): compressed from NVIDIA-Nemotron-Nano-9B-v2 with NVIDIA's Elastic framework. A hybrid Mamba2-Transformer with context up to 262K.
  • Nemotron 3 Ultra 550B-A55B (June 4, 2026): the largest, also LatentMoE with multi-token prediction, 550B total and 55B active, up to 1M tokens of context. License: OpenMDW 1.1.
  • Nemotron 3.5 Lightning 30B-A3B (August 11, 2026): a hybrid MoE with 30B total and 3B active parameters. The card lists up to 1M tokens, with 256K used for its single-H100 deployment, and says it ships alongside several speculative-decoding variants. License: OpenMDW 1.1.

The pattern across 2026: the same hybrid design at four scales, with newer releases moving to the OpenMDW 1.1 license.

Side lines

  • Nemotron-H: the earlier hybrid generation, with repositories from April and June 2025. Nemotron-H-56B-Base-8K is a completion model of mostly Mamba-2 and MLP layers with ten attention layers. Nemotron-H-8B-Reasoning-128K has four attention layers. These cards name the NVIDIA Internal Scientific Research and Development Model License, not the open model license, so check that it covers your use.
  • Nemotron Nano: the Nano 2 models, NVIDIA-Nemotron-Nano-9B-v2 (August 12, 2025) and 12B-v2 (August 21, 2025). Hybrid Mamba-2 and MLP with four and six attention layers respectively, under the NVIDIA Open Model License.
  • Nemotron Labs: experiments such as Nemotron-Labs-Diffusion-14B (April 2026) under the Nemotron Open Model License.
  • Nemotron Cascade: Nemotron-Cascade-2-30B-A3B (March 2026), which the card says supports up to a 1M-token context.
  • Nemotron Speech: nemotron-speech-streaming-en-0.6b (December 2025), a 0.6B-class streaming speech model under the NVIDIA Open Model License.
  • Nemotron Parse: NVIDIA-Nemotron-Parse-v1.1 (November 2025), a roughly 1B-parameter document parsing model.
  • Nemotron VoiceChat: NVIDIA-NemotronLabs-VoiceChat-11B (July 29, 2026), an 11B hybrid Mamba/Transformer under OpenMDW 1.1.

Which model to use today

These picks use sizes, context length and license only.

  • One mid-size GPU, agentic or reasoning chat: Nemotron 3.5 Lightning (30B total, 3B active) or Nemotron 3 Nano 30B-A3B. Lightning is newer and OpenMDW-licensed.
  • Smallest model: Nemotron 3 Nano 4B, at 4B and 262K context.
  • Very long inputs: Super (120B) or Ultra (550B), both listed up to 1M tokens. At 550B total parameters, Ultra is a multi-GPU model.
  • License sensitivity: Lightning, Ultra and VoiceChat use OpenMDW 1.1. The Nemotron-H cards use a research-focused NVIDIA license, so prefer the 3 series unless you need the older models.
  • Speech or document parsing: use the speech and parse hubs, not the language models.

Inference engines need Mamba-2 support for these models. Follow the engine versions on each card. You can rent GPUs for any of them on Aquanode GPUs.

Nemotron generations

Every Nemotron generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
Nemotron 319631M to 560.5B2.9 GBRTX 3070$0.088/hr

Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.

Other Nemotron lines

Side lines from NVIDIA: specialised or companion models outside the main Nemotron generations.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
Nemotron-H38.1B to 56.3B18.1 GBRTX A5000$0.176/hr
Nemotron Nano64.0B to 13.2B19.9 GBRTX A5000$0.176/hr
Nemotron Labs53.8B to 13.5B8.6 GBRTX 3060$0.110/hr
Nemotron Cascade131.6B70.6 GBA100$1.31/hr
Nemotron Speech1618M2.8 GBRTX 3070$0.088/hr
Nemotron Parse1957M4.3 GBRTX 3070$0.088/hr
Nemotron VoiceChat111.1B49.6 GBA100$1.31/hr

Run and fine-tune Nemotron

Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.

Nemotron compared with other models

Side-by-side VRAM, context length and license for models of a similar size.

Best models by task

Nemotron models appear on these ranked task pages.

Sources

Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.

Start ComfyUI on a rented GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.