AIPS BOOM

Nemotron-3-Nano-Omni

4 min read

Nemotron 3 Nano Omni

Type: Model / NIM microservice Tags: NVIDIA, Nemotron, omnimodal, multimodal, VLM, audio, video, image, document intelligence, agentic AI, NIM Related: Nemotron, NIM-for-Vision-Language-Models, NVIDIA-NIM, NVIDIA-Agent-Intelligence-Toolkit, NVIDIA-NemoClaw, NVIDIA-OpenShell, NVIDIA-NeMo, NeMo-Data-Designer, NeMo-AutoModel, NeMo-RL, NeMo-Megatron-Bridge, NIM-for-Image-OCR, NIM-for-Object-Detection, NIM-for-Cosmos-Reason, NVIDIA-AI-Blueprints, NVIDIA-AI-Q-Blueprint, NVIDIA-Video-Search-and-Summarization-Blueprint, TensorRT-LLM, vLLM, NVIDIA-Dynamo, NVIDIA-DGX-Spark, NVIDIA-DGX-Station Sources: https://docs.nvidia.com/nim/vision-language-models/latest/examples/nemotron-3-nano-omni-30b-a3b-reasoning/api.html; https://docs.nvidia.com/nim/vision-language-models/latest/support-matrix.html; https://docs.nvidia.com/nim/vision-language-models/latest/release-notes.html; https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning; https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/; https://developer.nvidia.com/blog/nvidia-nemotron-3-nano-omni-powers-multimodal-agent-reasoning-in-a-single-efficient-open-model Last Updated: 2026-04-29

Summary

Nemotron 3 Nano Omni is NVIDIA’s current open omnimodal reasoning model for agentic workflows that need a single perception model across text, images, video, audio, documents, charts, and graphical interfaces. The latest NIM-for-Vision-Language-Models docs introduce it in the 1.7.0 release stream and provide a self-hosted NIM container for the nvidia/nemotron-3-nano-omni-30b-a3b-reasoning model.

Detail

Purpose

Agentic systems often need to reason over screens, PDFs, audio, video, images, and text in the same task. Nemotron 3 Nano Omni is positioned as the perception and context sub-agent for those workflows, reducing the need to chain separate vision, speech, and language models before handing context to planners or execution agents.

Current NIM surface

  • Current VLM NIM release notes list Nemotron 3 Nano Omni as an initial release in NIM for VLMs 1.7.0.
  • The NIM API page launches repository nemotron-3-nano-omni-30b-a3b-reasoning with release tag 1.7.0-variant.
  • The NIM uses a specialized base container; current docs point readers to the NIM container-variant notes for that path.
  • The API follows the OpenAI Chat Completions pattern and accepts image, video, and audio content in user messages.
  • Media can be passed by public URL or base64 data. Current examples show image, video, audio, and audio-in-video request patterns.
  • The model supports reasoning with text and vision inputs. Thinking is enabled by default and the response separates reasoning from final content.
  • Request controls include chat_template_kwargs.enable_thinking=false, thinking_token_budget, media_io_kwargs for video frame sampling, and NIM_VIDEO_PRUNING_RATE for Efficient Video Sampling.

Hardware and deployment

  • Current support matrix lists NIM for VLM release 1.7.0 for this model.
  • Generic NIM support requires CUDA compute capability >= 9.0 and sufficient memory.
  • BF16 profiles require 80 GB total GPU memory and run on 1 or 2 GPUs, with listed support including B300, B200 SXM, GB300, GB200 SXM, GH200 SXM, H200, H100, RTX PRO 6000 Blackwell Server Edition, and L40S with 2 GPUs.
  • Current supported video codecs/formats include H264, H265, VP8, VP9, and FLV codecs with MP4, FLV, and 3GP containers. The API examples list GIF/JPG/JPEG/PNG images, MP4 video, and WAV/MP3/FLAC audio payloads.
  • Release notes call out a current limitation where requests with videos larger than 1 GB may fail with a NanoNemotronVLProcessor error.

Model context

NVIDIA’s launch and technical blogs describe Nemotron 3 Nano Omni as a 30B-A3B hybrid mixture-of-experts model with Conv3D video processing, Efficient Video Sampling, and a long-context omnimodal design. The model is part of the broader Nemotron 3 family and is intended to work with other Nemotron reasoning models, proprietary planners, NVIDIA-Agent-Intelligence-Toolkit, NVIDIA-NemoClaw, and NVIDIA-OpenShell agent runtimes.

Use this page for NVIDIA Nemotron 3 Nano Omni itself. Use NIM-for-Vision-Language-Models for the broader VLM NIM family, Nemotron for the whole Nemotron model family, Nemotron-Parse for the document-parsing VLM, and NIM-for-Image-OCR or NIM-for-Object-Detection for the NeMo Retriever extraction microservices.

Connections

Source Excerpts

  • “initial release of NVIDIA Nemotron 3 Nano Omni”
  • “accepts images, videos, and audio as input”
  • “30B-A3B hybrid MoE”

Resources

Graph View

Backlinks