← Back to feed

NVIDIA shipped Nemotron 3 Nano Omni on April 28, and the architecture decision is the story. Where competing open multimodal systems chain separate vision, audio, and language models — passing data across model boundaries with latency and context loss at each hop — Nemotron 3 Nano Omni folds all three modalities into one 30B-parameter hybrid MoE system with 3B active parameters at inference. The result is an open omni model that NVIDIA claims runs 9.2x more concurrent agents per GPU at equivalent interactivity thresholds.

Architecture

The model uses a hybrid Mamba-Transformer design — 23 mixture-of-experts layers with 128 experts and six grouped-query attention layers — combined with dedicated vision and audio encoders. The active parameter count stays at 3B regardless of input modality. Context length is 256,000 tokens. On Blackwell GPUs with NVFP4 quantization, NVIDIA says the throughput advantage extends to 7.4x for multi-document reasoning workloads.

The 30B-A3B architecture is a direct descendant of the text-only Nemotron 3 Nano announced earlier this year. Omni extends it by integrating multimodal encoders natively into the model loop rather than as preprocessing stages.

Benchmark Position

On the Arena-Hard-V2 composite (hard prompts + creative writing), Nemotron 3 Nano Omni scores 67.65 — ahead of Qwen3-30B-A3B-Thinking-2507 at 57.80 and GPT-OSS-20B at 48.55. On math benchmarks, MATH accuracy reaches 82.88 versus Qwen3’s 61.14 at the same parameter tier.

Six leaderboard top spots claimed at launch: MMlongbench-Doc, OCRBenchV2, WorldSense, DailyOmni, VoiceBench, and MediaPerf (video understanding efficiency). RULER-100 long-context retrieval: 92.92% at 256K, 91.25% at 512K, 86.34% at 1M tokens — outperforming GPT-OSS-20B (89.40% at 256K) and Qwen3-30B (no 512K/1M entries).

SWE-Bench via OpenHands scaffold: 38.76%. That number uses a different runner from the mini-SWE-agent standard and isn’t directly comparable to the 70-80%+ scores posted by frontier reasoning models.

Adoption and Distribution

H Company’s computer-use agent is running Nemotron 3 Nano Omni as its perception backbone, using the model’s native 1920x1080 input resolution for GUI reasoning. Palantir, Foxconn, and Eka Care are in production. Dell, Docusign, Infosys, and Oracle are listed as evaluating.

The model is free on OpenRouter (nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free), available on Hugging Face, and deployable via NVIDIA NIM microservices on build.nvidia.com. NVIDIA is releasing full weights, training recipes, and all data for which it holds redistribution rights.

What This Signals

The Nemotron 3 family — Nano, Super, Ultra — is accumulating 50 million downloads in the past year across text-only versions. Adding the Omni variant extends NVIDIA into the open multimodal category, where the current competition is primarily Qwen’s omni line and smaller open projects. The efficiency numbers are the pitch: if NVIDIA can deliver comparable accuracy at 3x-9x the concurrency per GPU, the compute economics argument for enterprise agent deployments shifts.

Super and Ultra releases are described as coming in the next months.