New Nvidia Nemotron 3 Nano Omni Update Changes Everything!

The video introduces Nvidia's Nemotron 3 Nano Omni, released April 28th, 2026, as a major leap in open-source multimodal AI. Unlike most AI models that specialize in a single modality, this model can process text, images, audio (up to 1 hour), and video (up to 2 minutes) simultaneously in a single pass, with a 256K context window for handling large documents. The presenter frames this as transformative for small business owners who are overwhelmed by PDFs, voice notes, screen recordings, and training videos.

The technical architecture is explained in accessible terms. The model uses 30 billion parameters but only activates roughly 3 billion at a time through a Mixture of Experts (MoE) design, where specialized sub-models are selectively engaged depending on the query. For video processing, Nvidia introduced Conv3D Tubelet Embedding, which processes two video frames simultaneously instead of one, and Efficient Video Sampling, which skips redundant frames where little is happening and focuses attention on frames with meaningful activity.

Benchmark results are presented across multiple evaluation categories: OCR Bench V2 (65.8% vs. 61.2% for the prior model), Video MME (72.2% vs. 70.5%), Voice Bench (89.4% vs. 88.8%), M Long Bench Doc for document analysis (57.5%), and Screen Spot Pro for on-screen UI understanding (57.8%). Nvidia claims 9.2x efficiency on video tasks and 7.4x on multi-document tasks, meaning a task that previously took 9 minutes now takes roughly 1 minute.

The presenter outlines deployment options: DeepInfra for API-based access with OpenAI-compatible endpoints, and Hugging Face for local deployment via Unsloth, with multiple quantization options (BF16, FP4, NVFP4) for varying hardware capabilities. A practical use case is illustrated with a real estate agent using the model to auto-generate property descriptions and identify issues from 50 walkthrough videos. The model's Screen Spot Pro score is highlighted as enabling agentic screen interaction — bots that can read a screen and autonomously click, fill forms, or gather data.

The video closes with a broader observation that open multimodal AI has advanced dramatically in one year, and promotes two communities: the paid AI Profit Boardroom and the free AI Success Lab with 67,000 members.