Can you run it?
This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Install & run
Download the ungated safetensors from the verified nvidia org on Hugging Face. The repo id has
no "NVIDIA-" prefix (nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16). Serve with vLLM,
SGLang or NVIDIA NIM; a TensorRT-LLM deploy cookbook is linked on the card. Pin the exact
revision and verify checksums before loading.
Hardware & VRAM requirements
Small and portable: 31B total / ~3B active runs on modest single-GPU hardware. Official BF16 (62 GB), FP8 (33 GB) and NVFP4 (21 GB) variants are published, and the card gives minimum GPUs per precision: one H100 80GB for BF16, one L40S 48GB for FP8, one RTX 5090 32GB (or DGX Spark, Jetson Thor) for NVFP4. Budget additionally for the multimodal input pipeline (image, audio and video preprocessing).
Serving stacks
vLLM, SGLang and NVIDIA's own NIM self-hosted microservices, plus a linked TensorRT-LLM deploy cookbook. NIM runs on your own infrastructure, so data stays in your enclave. The multimodal serving path is less mature than the text-only members of the family.
Safe-deployment controls & Deployment Ceiling
Deployment Ceiling: T2 (bounded), conditional. The multimodal input surface widens the misuse path and there is no model-level safety evaluation, so the controls are the deployer's responsibility. They are the multimodal-capable Nemotron-3-Content-Safety guard classifier (text and images), the NeMo Guardrails toolkit and the garak red-team scanner.
Red-team the video/audio/image input path for the specific use case before any autonomous use. The irrevocable licence imposes no field-of-use restriction. Complete the pre-deployment gate in the safe-deployment playbook.
Available quantizations
Official NVIDIA BF16 (62 GB), FP8 (33 GB) and NVFP4 (21 GB) variants are published; the FP8 and NVFP4 repos date from 2026-04-24. Check engine support per precision. The card's SGLang section covers the BF16 variant only ("NVFP4 and FP8 support to come"), while its vLLM examples use the NVFP4 and FP8 repos.
Fine-tuning & adaptation
A training recipe exists in the NeMo repo (docs/nemotron/omni3, SFT then RL), so adaptation is
realistic. Reproduction is not: the recipe surfaces only the public subset of the alignment
corpus, and the full upstream corpus of 20 RL datasets / 25 environments / ~2.3M rollouts is not
open. Adapt from the released checkpoint and do not expect a full re-run.
API / OpenAI-compatible integration
Serve behind vLLM/SGLang or NVIDIA NIM for OpenAI-compatible endpoints; the multimodal input path
may need the documented preprocessing. A hosted build.nvidia.com preview API exists for
evaluation only. It reserves use of content "to improve NVIDIA products and services, including
AI models".
How this scores
The ownership factors this domain covers, drawn from the one entry record.
ReliabilityIs it reliable and good enough for the job?
ModerateOperational is 3: the model runs on mainstream stacks with official BF16 / FP8 / NVFP4 variants, but the multimodal serving path is narrower, with SGLang supporting only the BF16 variant. Safety is 3: there is no model-level safety evaluation over a wider multimodal misuse surface. Under the ownership rule, a safety score of 3 caps this factor at moderate.
Doesn't extract your dataDoes running it keep your knowledge and data yours?
StrongSelf-hosted via NIM (or the linked TensorRT-LLM path), the weights and inference run on your own infrastructure and data never leaves your enclave. The NVIDIA Open Model Agreement claws back no rights: the grant is explicitly irrevocable and terminates only on the licensee's own IP litigation. That is the strong case for a self-hostable model.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.