Alpha
Contact
Assess · NVIDIA Nemotron 3 Nano Omni

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 60/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

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. It is intended for small, portable, multimodal reasoning and agentic work where an open, permissively licensed model is wanted.

Autonomous deployment without a self-run red-team of the multimodal input path is out of scope, 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

The openness posture is weaker than the rest of the family. 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, and the model is classified open_weights, not open_weights_recipe.

There is no model-level safety evaluation. Safety is post-training data curation only, over a wider multimodal misuse surface.

A recipe exists in the NeMo repo, but it surfaces only the public subset of the alignment corpus, so full reproduction is not possible.

The NVIDIA Open Model Agreement's version line is "v. March 9, 2026" in both the HTML and the PDF, but the web page header shows a release date of April 2, 2026. The operative text is the same either way.

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). The weights are ungated safetensors, with a model card and a clear licence. 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 training recipe exists 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 repo recipe alone does not raise the tier.

License terms & permitted use

Nano Omni is governed by the NVIDIA Open Model Agreement (v. March 9, 2026), a third distinct family licence whose text is near-identical to the NVIDIA Nemotron Open Model License. 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". Commercial use is allowed, and the Trustworthy-AI, acceptable-use and guardrail clauses are absent.

The licence terminates only if the licensee brings patent or copyright litigation over the Work or an output from it. It is source-available and not OSI. The text is identical in the HTML and the PDF and matches the Nemotron Open Model License apart from the attribution notice. Both formats end "(v. March 9, 2026)"; the April 2, 2026 date appears only in the HTML header and the PDF file name.

Section 7 ends with a general indemnity ("You will indemnify and hold harmless NVIDIA from and against any claim by any third party arising out of or related to your use or distribution of the Works"). It is a liability term and not a use restriction. These terms underlie the legal score of 3.

Supply-chain & provenance

Weights are distributed from the verified nvidia org on Hugging Face as ungated safetensors with an official BF16 variant. 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 are not verified, 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 and 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 reach the open-source exemption cleanly.

Training compute is undisclosed, but at 31B / ~3B active on ~717B tokens a systemic-risk designation is not indicated. An EU deployer must also weigh the multimodal input surface.

Benchmarks & evaluation

Nano Omni is a small multimodal reasoning model, competitive for its class and modality. OneHill has not run its own benchmarks, and no independent third-party re-runs of the multimodal benchmarks are aggregated, so the figures are publisher and architectural (the model card). The item 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. 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 holds the safety dimension at 3 and not 2, while the missing model-level evaluation over the wider surface holds it below 4. The guard-classifier scores are NVIDIA self-reported.

The model's safety.md subcard says safety evaluations were run with AEGIS v2, Garak, RTVLM and VLGuard plus custom child-safety and dehumanization datasets. It describes an AEGIS 1.0 gate of at least 95 percent with zero sexual-minor items, runs with reasoning on and off, and an over-refusal evaluation. It publishes no scores, per-category results or residual-risk disclosure, and NVIDIA ran it. This is a self-reported process statement, not a published evaluation, and a published result table would justify a higher safety score.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Moderate

The NVIDIA Open Model Agreement is permissive: perpetual, royalty-free and irrevocable, with commercial use allowed, no acceptable-use or field-of-use restriction, ungated weights and fine-tuning permitted. It is held at moderate, not strong. The licence is Apache-2.0-derived but non-OSI, and termination reaches an output from the Work. It is a distinct third family licence whose version line (March 9, 2026) differs from the April 2, 2026 release date on its web page. The released recipe surfaces only the public subset of the alignment corpus. It is an open_weight community licence with conditions, short of Ultra's unconditional OpenMDW grant.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyOpen weights, a model card and a clear licence, but the openness posture is weaker than the rest of the family.
Legal3/5how permissive and clean the licence is for real commercial useThe NVIDIA Open Model Agreement is a commercially permissive grant: Apache-2.0-derived, perpetual, royalty-free and irrevocable, with commercial use allowed.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

open_weights, not open_weights_recipe. The weights are ungated and inspectable and there is a model card and licence, but the card does not carry the family's open-data/recipe self-description, links only an image-training dataset, and the recipe that exists surfaces only the public subset. The weights are visible but the data and full recipe are not clearly released, which is the anchor for moderate, not strong.

How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelThe verified nvidia org on Hugging Face distributes ungated safetensors with official BF16 / FP8 / NVFP4 variants and a clear canonical source (checklist ~4/8).
Governance3/5how accountable and well-documented the publisher isNVIDIA is a reputable, accountable publisher with a verified hub presence, an active release cadence and a general corporate security-reporting path.
What this means for adoptionNano Omni gives you a small, portable, natively multimodal model under a clean irrevocable licence, and self-hosting keeps your data and rights yours, so data control is strong. Ownership stays partial for three reasons. Use-and-modify is moderate (a non-OSI third licence and a public-subset-only recipe). Transparency is moderate (open_weights, not open_weights_recipe: the card does not release the data or advertise the recipe). Reliability is moderate (no model-level safety evaluation over a wider multimodal misuse surface). Self-host, assemble the multimodal guard stack, red-team the video/audio/image input path, and keep a copy of the licence version the weights were downloaded under.

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.

Model cardread2026-11-19
The Nemotron 3 Nano Omni model card is on the verified nvidia HF org (nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, no "NVIDIA-" prefix).
Licenceread2026-11-19
The NVIDIA Open Model Agreement (v.
Documentationread2026-08-03
A training recipe for Nano Omni exists in the NeMo repo (github.com/NVIDIA-NeMo/Nemotron docs/nemotron/omni3, SFT then RL).
Model cardread2026-11-19
There is no dedicated model-level safety evaluation for Nano Omni; safety is post-training data curation only, over a wider multimodal (video/audio/image) misuse surface.
Documentationread2026-11-19
The nvidia org on Hugging Face is verified, and the Nano Omni weight repo is ungated safetensors with official BF16, FP8 and NVFP4 variants.
Terms of serviceread2026-08-03
NVIDIA NIM is self-hosted deployment tooling ("data never leaves your secure enclave"), and the Nano Omni card links a TensorRT-LLM deploy cookbook, so self-hosting supports data control and NVIDIA is not an AOI inference-provider entry.
Third-party analysisread2026-08-03
NVIDIA is not a signatory to the EU GPAI Code of Practice (EU signatory list), and no NVIDIA training-content summary or copyright policy for Nemotron is published.