Nemotron 3 Nano Omni
Open 30B-A3B hybrid MoE that reasons over text, image, audio and video in one loop for agents.
About
Nemotron 3 Nano Omni is NVIDIA's open omni-modal model, released April 28, 2026. It is a 30-billion-parameter hybrid mixture-of-experts with about 3 billion active parameters per token, and it reasons over text, images, audio, and video in one loop instead of chaining separate encoders. It ships under the NVIDIA Open Model License with weights, datasets, and training recipes published, and it runs on a single H100.
NVIDIA claims best-in-class scores on document, video, and speech leaderboards (MMlongbench-Doc, WorldSense, DailyOmni, VoiceBench) and roughly 9x the throughput of Qwen3-Omni on B200 GPUs. Third-party April 2026 evaluations put it at about 71 on MMLU-Pro, 76 on HumanEval+, and 73 on MMMU, behind the closed flagships on text but ahead on multimodal. It is meant as the perception layer for agentic systems: computer-use agents, document intelligence, and customer-service agents that watch and listen.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- MMMU: ~73 (third-party, Apr 2026)
- MMLU-Pro: ~71 (third-party, Apr 2026)
- HumanEval+: ~76 (third-party, Apr 2026)
- Leads open-model leaderboards on OCRBenchV2, MMLongBench-Doc, Video-MME, WorldSense, VoiceBench, per NVIDIA
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 50M+ downloads across the Nemotron 3 family in a year per NVIDIA with enterprise early adopters; active NIM and Hugging Face distribution. Fully open weights, datasets, and recipes. Release facts corroborated across NVIDIA developer forums and third-party evals; docked for NVIDIA-reported leaderboard claims and third-party text benchmarks trailing closed flagships.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 6: 50M+ downloads across the Nemotron 3 family in a year per NVIDIA; early adopters include Aible, Foxconn, Palantir, H Company
- · Maintenance 8: released Apr 28, 2026; active NIM/Hugging Face distribution with FP8 and NVFP4 variants
- · Openness 10: open weights, datasets, and training recipes under NVIDIA Open Model License (0.7*10) + public Hugging Face/NIM distribution (0.3*10)
- · Manual 8: release facts (30B-A3B hybrid MoE; 256K context; single-H100 capable) corroborated across NVIDIA developer forums, Pulse2, and third-party evals; docked for NVIDIA-reported leaderboard claims and third-party text benchmarks trailing closed flagships