Production model

Nemotron 3 Nano 30B-A3B

NVIDIA · Released 2025-12-15 · 30B parameters (3.5B active) · 1,000,000 ctx · open weights

Architecture

Positional encoding	—
Norm placement	—
Norm type	—
QK-Norm	—
Activation	—
Attention	Hybrid: 23 Mamba-2 + 6 GQA attention layers (2 KV groups); MoE FFN on 23 layers
MoE	Sparse MoE — 128 routed experts + 1 shared expert per MoE layer, top-6 routing
Other	52 layers total: 23 Mamba-2, 6 attention, 23 MoE FFN, 30B total / 3.5B active per token, 1M-token native context (default 256K due to VRAM), Reasoning + non-reasoning unified in one model with thinking-budget control, NVIDIA Open Model License

Techniques used

Grouped-Query Attention — GQA with 2 KV groups on the 6 attention layers; the other 23 sequence-mixing layers are Mamba-2.
Linear Attention — 23 Mamba-2 layers (the descendant of structured-state-space variants of the linear-attention family) in a 52-layer hybrid stack.
Sparsely-Gated MoE — 128 routed experts + 1 shared expert per MoE layer; top-6 routing; 23 MoE layers in the 52-layer hybrid stack.

Nemotron 3 Nano 30B-A3B (December 2025) is NVIDIA’s first hybrid Mamba-Transformer MoE release in the Nemotron 3 family. The model card documents an unusual 52-layer composition:

23 Mamba-2 state-space layers. Linear-cost sequence mixing — the bulk of the layer budget — providing $O(L)$ scaling for the 1M-token context.
6 GQA attention layers with 2 KV groups. The few full-attention layers anchor high-fidelity token-token reasoning that Mamba-2 alone underperforms on.
23 MoE FFN layers. 128 routed experts plus one always-on shared expert, top-6 routing.

Active parameters per token: 3.5B (out of 30B total).

The model supports a configurable thinking budget: a single unified model produces internal reasoning traces when asked or responds directly otherwise. Native context is 1M tokens; the default deployment config caps at 256K due to VRAM.

Per the citation policy, architectural slots not surfaced in the model card or arXiv white paper — norm type and placement, activation function, positional encoding for the attention layers, QK-Norm presence — are left null rather than guessed.

The Nemotron 3 family white paper (arXiv:2512.20856) announces Super and Ultra variants forthcoming after the Nano release. Those add LatentMoE and Multi-Token Prediction layers per the paper but their weights are not yet released at this entry’s verified date.

For NVIDIA’s earlier dense flagship see Nemotron-4 340B (a 2024 release; not present in this knowledge base as a separate entry).

Sources

Export

BibTeX

@article{arxiv_2512_20856,
  title         = {Nemotron 3 Nano 30B-A3B},
  author        = {NVIDIA},
  year          = {2025},
  eprint        = {2512.20856},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2512.20856}
}

CSL JSON

{
  "id": "arxiv_2512_20856",
  "type": "article-journal",
  "title": "Nemotron 3 Nano 30B-A3B",
  "author": [
    {
      "literal": "NVIDIA"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2025
      ]
    ]
  },
  "URL": "https://arxiv.org/abs/2512.20856",
  "number": "2512.20856",
  "source": "arXiv"
}

RIS

TY  - JOUR
TI  - Nemotron 3 Nano 30B-A3B
AU  - NVIDIA
PY  - 2025
JO  - arXiv
AN  - arXiv:2512.20856
UR  - https://arxiv.org/abs/2512.20856
ER  -

Models like this

Mistral Large 3 (675B)Mistral AI33% overlap
Qwen3 30B-A3BAlibaba (Qwen Team)29% overlap
Qwen3 235B-A22BAlibaba (Qwen Team)25% overlap