Model

Nemotron 3 Nano Omni

OPEN

NVIDIA · Nemotron 3 family · released Apr 29, 2026

First truly omni-modal Nemotron 3 model (text/vision/audio/video natively unified, not vision-language with audio bolted on); smallest member of the family alongside Nemotron 3 Super and Ultra (already in corpus).

ReasoningCodingVisionFunction callingTool useAgentic

1485.6

Elo · rank #139

Parameters

30B

Active params

3B (MoE)

Context

256K tokens

Architecture

Hybrid Mamba-2/Transformer MoE, omni-modal (23 Mamba-2 layers + 23 MoE layers of 128 experts + 6 GQA layers), 30B total / 3B active

License

NVIDIA Open Model License

Languages

—

API price (in/out)

$0.075 / $0.3

Modalities

text · vision · audio · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision88.5%#24

best: Molmo 72B · 96.3%

AIMEMath82.1%#71

best: GPT-5.2 · 100.0%

ChartQAVision90.3%#4

best: MiniMax-VL-01 · 91.7%

CharXivVision63.6%#34

best: Qwen3.8-Flash-Next · 90.6%

DocVQAVision95.6%#4

best: Qwen2-VL-72B · 96.5%

GPQA DiamondReasoning72.2%#121

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning5.0%#127

best: Claude Opus 5 · 64.7%

IFBenchReasoning74.2%#22

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding63.2%#95

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVistaVision82.8%#12

best: Seed 2.1 Pro · 90.7%

MMLU-ProKnowledge77.3%#69

best: Claude Fable 5 · 91.5%

MMMU-ProVision53.0%#47

best: Claude Opus 4.7 · 85.5%

MMMUVision70.8%#50

best: Claude Fable 5 · 89.3%

OCRBenchVision87#19

best: InternVL3-78B · 906

OSWorldAgents47.4%#2

best: Seed 2.1 Pro · 78.8%

RefCOCOVision90.5%#10

best: Qwen3.5-Omni-Plus · 95.0%

τ²-Bench TelecomAgents42.7%#34

best: Claude Opus 4.6 · 99.3%

TextVQAVision81.0%#11

best: Molmo 2 8B · 85.7%

Video-MMEVision72.2%#15

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Conditional / custom

QLoRA20.9 GB1× RTX 3090 24GB

LoRA64.4 GB1× A100 80GB

Full fine-tune482.0 GB4× H200 141GB

QLoRA SFT on ~10k samples ≈ $15.45 (1× RTX 3090 24GB)

Plan a fine-tune →

Nemotron 3 family

Elo progression across releases

Nemotron 3 SuperMar 20261805.3 Nemotron 3 Nano OmniApr 20261485.6 Nemotron 3 Ultra 550B A55BJun 20262155.8

Resources

HFnvidia/Nemotron-3-Nano-Omni-30B-A3B DOCShttps://developer.nvidia.com/blog/nvidia-nemotron-3-nano-omni-powers-multimodal-agent-reasoning-in-a-single-efficient-open-model/APIAvailable via NVIDIA API

API price $0.075/$0.3 · each benchmark row carries its own source badge (see methodology)