Nemotron 3 Nano Omni 30B A3B Reasoning

NVIDIA released the open Nemotron 3 Nano Omni on April 28, 2026, combining vision, audio, and language into a single model with a 30B-A3B hybrid Mixture-of-Experts architecture that activates only 3B parameters per token. NVIDIA claims up to 9x higher throughput than other open omni models at equivalent interactivity, positioning it as the perception sub-agent paired with higher-tier Nemotron 3 Super and Ultra reasoners. On a single H200 with 8K input and 16K output, the model delivers 3.3x the inference throughput of Qwen3-30B-A3B and 2.2x that of GPT-OSS-20B, per NVIDIA's developer blog. On MediaPerf video tagging it processes 9.91 hours of video per hour versus roughly 3.8 h/h for Qwen3-VL. NVIDIA published weights, datasets, and training techniques, with NVFP4 quantization, NeMo tooling, and on-prem deployment supported across over 50 early partners.

DeepInfra became an official launch partner for NVIDIA Nemotron 3 Nano Omni on April 28, 2026, making the first multimodal entry in the Nemotron 3 family available from day one. The open model processes images, video, audio, documents, and text in a single inference pass, with NVIDIA claiming 9x higher throughput than other open omni models at the same interactivity level. The model uses a hybrid Mixture of Experts and Mamba-Transformer backbone extended from Nemotron 3 Nano, with 3D convolution layers and Efficient Video Sampling for low-cost video reasoning. A hybrid MoE design activates 3B of 30B total parameters per token, enabling developers to build always-on multimodal sub-agents for computer use and audio-video understanding with minimal code.

NVIDIA announced the Nemotron 3 Nano Omni on April 28, 2026, as an open omnimodal inference model that unifies vision, speech, and language processing into a single system, eliminating the latency and context loss of chained separate models. The model targets agent workflows such as computer use, document analysis, and speech/video reasoning. NVIDIA reports the Nemotron 3 Nano Omni ranks first in six categories spanning complex document recognition, video, and audio understanding benchmarks. It significantly outperforms its predecessor, the Nemotron Nano VL V2, across the board, with particularly substantial gains on OSWorld, a benchmark for multimodal agents measuring real-world computer task completion.