Can you own it?

Intended & out-of-scope use

Nemotron 3 Nano Omni is the multimodal member of the family: 31B total / roughly 3B active, a Mamba2-Transformer hybrid Mixture-of-Experts that takes video, audio, image and text in and returns text, with a 256k-token context. Its intended use is small, portable, multimodal reasoning and agentic work where you want an open, permissively-licensed model.

Out of scope: autonomous deployment without a self-run red-team of the multimodal input path, because there is no model-level safety evaluation and the video/audio/image surface widens the misuse path. Ordinary commercial use and modification are unencumbered - the licence is irrevocable and carries no field-of-use restriction.

Known limitations, bias & failure modes

• Weaker openness posture. Unlike the Super/Nano/Ultra cards, the Nano Omni card does not carry the "open weights, training data, and recipes" self-description; it links only an image-training dataset and a deploy cookbook. It is classified open_weights, not open_weights_recipe.
• No model-level safety evaluation. Safety is post-training data curation only, over a wider multimodal misuse surface.
• Recipe surfaces only the public subset. A recipe exists in the NeMo repo, but it surfaces only the public subset of the alignment corpus, so full reproduction is not possible.
• Licence date inconsistency. The NVIDIA Open Model Agreement's on-document date (v. 2026-03-09) is inconsistent with a later HTML date and is unverified against the PDF.

The offsetting advantage is a small, portable, natively multimodal model under a clean irrevocable licence.

Openness tier & components

Nano Omni is open_weights (tier 3). Weights are ungated safetensors, with a model card and a clear licence, but the card does not carry the family's open-data/recipe self-description and links only an image-training dataset and a TensorRT-LLM deploy cookbook. A genuine training recipe does exist in the NeMo repo (docs/nemotron/omni3, SFT then RL), but it surfaces only the public subset of the alignment corpus, and the card signals do not carry the recipe claim. On card signals this is open_weights, not open_weights_recipe: the openness tier is not forced up on the repo recipe alone.

License terms & permitted use

Nano Omni is governed by the NVIDIA Open Model Agreement (v. 2026-03-09), a third distinct family licence whose text is near-identical to the NVIDIA Nemotron Open Model License. On the corrected reading it is Apache-2.0-derived and permissive: a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable license to reproduce, prepare Derivative Works of ... sublicense, and distribute the Work ... in source or object form", with commercial use allowed and the Trustworthy-AI, acceptable-use and guardrail clauses confirmed absent. It terminates only on the licensee bringing patent or copyright litigation over the Work or an output from it. It is source-available and not OSI. The on-document date is inconsistent with the HTML date and unverified against the PDF, so confirm the version before relying on it. These are the load-bearing inputs to the legal (3) score.

Supply-chain & provenance

Weights are distributed from the verified nvidia org on Hugging Face as ungated safetensors with an official BF16 variant. Note the repo id (nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16) has no "NVIDIA-" prefix, unlike its siblings, so pin the exact path. The checkpoint trust checklist scores about 4/8: published checksums, checkpoint scanning and a signing/attestation policy were not verified this pass, which is why provenance is a 3. Pin the exact revision and verify checksums.

EU AI Act posture

Nano Omni is GPAI. The licence is irrevocable with no acceptable-use restriction, but there is no formal EU GPAI documentation package, no copyright policy, and NVIDIA is not a Code of Practice signatory; the open-data posture is weaker than the rest of the family. A non-OSI licence with a weaker open-data posture most likely does not cleanly reach the open-source exemption. Training compute is undisclosed, but at 31B / ~3B active on ~717B tokens a systemic-risk designation is not indicated. An EU deployer must additionally weigh the multimodal input surface.

Benchmarks & evaluation

Nano Omni is a small multimodal reasoning model competitive for its class and modality. OneHill did not run its own benchmarks this pass, and no independent third-party re-runs of the multimodal benchmarks were gathered, so the figures are publisher and architectural (the model card). This is marked partial and holds the performance dimension at 3.

Independent safety evaluation

There is no dedicated model-level safety evaluation for Nano Omni, and its multimodal input surface (video/audio/image) widens the misuse path without a published characterisation of it. Safety is post-training data curation only. The Nemotron-3-Content-Safety classifier is multimodal (text and images) and so is partly attributable, and NeMo Guardrails and garak are Apache-2.0 - that downloadable stack is what holds the safety dimension at 3 rather than 2, while the missing model-level evaluation over the wider surface holds it below 4. The guard-classifier scores are NVIDIA self-reported.