NVIDIA / HYBRID OMNI / 2026

Nemotron 3 Nano Omni

TOTAL33.0B pkgparameters

ACTIVE≈3Bper token

DEPTH52 + encodersrepeated blocks

HIDDEN2,688main trunk

CONTEXT128Knative tokens

CHECKPOINT66.03GB17 shards · NVIDIA Open Model

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 1.11P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KIMAGE · VIDEO · AUDIO · TEXT

TRANSFORMER STACK52 HYBRID LAYERS23 Mamba · 23 MoE · 6 GQA · 128 experts top-6 + shared

hybrid residual sequence · state-space and attention layers alternate roles

OUTPUT1.0KTOKENS

wraps every repeated block

MAMBA ⇄ GQA · CONV3D + EVSConv3D 将相邻两帧融合为 tubelet，EVS 再丢弃后续帧的静态视觉 token，控制视频序列长度。

Nemotron MoE · 30B BACKBONE≈3B active path · total capacity ≠ active compute

BACKBONE52Lhybrid

MAMBA23state-space

MOE23128 → 6 + 1

GQA632Q / 2KV

VISION1K–13Kpatches/image

CONTEXT128Kcheckpoint