NVIDIA Nemotron 3 Nano Omni — Spécifications
Rédigé par Mustafa Ihsan d’après la documentation officielle du fournisseur · Dernière mise à jour
| Développeur | NVIDIA |
|---|---|
| Type | Multimodal (omni) |
| Modalité | Texte, image, audio, vidéo → Texte |
| Paramètres | 30 milliards au total / ~3 milliards actifs (MoE) |
| Fenêtre de contexte | 256 K |
| Licence | Accord open source NVIDIA (NVIDIA Open Model Agreement) |
| Poids ouverts | Oui |
| Publié | 2026 |
| Fournisseurs d'API | Hugging Face, OpenRouter, NVIDIA NIM |
Exécutez-le localement
| VRAM (FP16/BF16) | ~62 Go |
|---|---|
| VRAM (4 bits) | ~21 Go (NVFP4) |
| GPU minimal | RTX 5090 32 Go (NVFP4) / H100 80 Go (BF16) |
Résultats de référence
| OCRBench V2 | 67.04 |
|---|---|
| Video-MME | 72.2 |
| OSWorld | 47.4 |
| Speech IF | 89.39 |
What is NVIDIA Nemotron 3 Nano Omni?
NVIDIA Nemotron 3 Nano Omni is an open omni-modal model: it sees, hears, watches and reads
— text, image, audio and video in, text out — from a single 30B-A3B mixture-of-experts that
activates only about 3B parameters per token. It is a Mamba-Transformer hybrid with a 256K
context, released under the NVIDIA Open Model Agreement, which permits commercial use. It
scores 67.04 on OCRBench V2, 72.2 on Video-MME, 47.4 on OSWorld and 89.39 on Speech IF.
Four input modalities in one 30B model that runs on a single RTX 5090 at NVFP4 (about
21 GB) is the notable part. The usual way to build an application that handles audio, video,
images and text is to chain three or four specialist models, each with its own deployment,
latency budget and failure mode. Nemotron 3 Nano Omni collapses that into one, which is why
its 47.4 on OSWorld — a computer-use benchmark that requires seeing a screen and acting on it
— is more interesting than the raw number suggests. There is no public per-token API price;
the cost is the hardware, which for a single-GPU deployment is unusually approachable for a
model of this breadth.
NVIDIA Nemotron 3 Nano Omni pricing
NVIDIA Nemotron 3 Nano Omni has no public per-token API price. It is open-weight, so the only cost is the hardware you run it on — see the local-hardware table above and size a GPU with the Calculateur de VRAM.