Nemotron 3 Nano Omni

Claims and evidence

• Released 28 April 2026 ● Primary source[1]At a Glance: Availability [2]Release Date [3]

• Status AvailableNVIDIA has no deprecation page for its open models; the weights were still downloadable from the NVIDIA organisation on Hugging Face on 2026-10-01. ● Primary source[2]

• Derived from (other) Nemotron 3 NanoUses the Nemotron 3 Nano 30B-A3B language model as its backbone, combined with the C-RADIOv4-H vision encoder and a Parakeet speech encoder. ● Primary source[2]Model Architecture: Network Architecture
• Successor of Nemotron Nano 2 VLThe developer blog compares its multimodal accuracy with the previous Nemotron Nano VL V2 model. ● Primary source[3]Figure 3

• Change · Modality Accepts audio input through a Parakeet speech encoder, alongside video, image and text input.Compared with Nemotron Nano 2 VL ● Primary source[2]Input(s)
• Change · Architecture Uses the Nemotron 3 Nano 30B-A3B hybrid mixture-of-experts language model as its backbone, with the C-RADIOv4-H vision encoder.Compared with Nemotron Nano 2 VL ● Primary source[2]Model Architecture: Network Architecture
• Change · Context length Context window of up to 256K tokens.Compared with Nemotron Nano 2 VL ● Primary source[2]At a Glance: Max context
• Change · Licensing Released under the NVIDIA Open Model Agreement.Compared with Nemotron Nano 2 VL ● Primary source[2]License/Terms of Use

• Input text, image, audio, video ● Primary source[2]Input(s) [1]At a Glance
• Output text ● Primary source[2]Output(s)
• Feature Reasoning modeOn by default; can be toggled. ● Primary source[2]At a Glance: Reasoning mode
• Feature Tool useThe card says the model supports tool calling. ● Primary source[2]Output(s)
• Open weights YesReleased under the NVIDIA Open Model Agreement in BF16, FP8 and NVFP4. ● Primary source[1]Open and Customizable, Deployable Anywhere [2]License/Terms of Use; Download Model Weights
• Context window 256K tokensThe model card writes 256k tokens. ● Primary source[1]At a Glance: Architecture [2]At a Glance: Max context
• Parameters 31B (~3B active per token)NVIDIA names the architecture 30B-A3B in the blogs; the model card gives 31B total parameters (3.1 x 10^10) with about 3B active per token. ● Primary source[2]At a Glance; Model Architecture
• Access Open-weights download, API, cloud partnerHugging Face, OpenRouter, build.nvidia.com as an NVIDIA NIM microservice and partner platforms. ● Primary source[1]At a Glance: Availability; Open and Customizable, Deployable Anywhere

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from NVIDIA on AI Radar →