NVIDIANVIDIA-Certified Professional: Agentic AI2026-109.5 Ensure compliance with licensing and regulatory standards.
Check permissions for every shipped artifact
Completed by the knowledge check
Learning objective
Map model, dataset, and software licenses to the permitted use.
Core idea
Before you select an agent’s components for production, match each artifact’s license to the operation you intend to perform. Downloading weights, fine-tuning them, running an internal service, distributing an appliance, and sharing a dataset are different operations. Permission for one does not establish permission for all of them.
Your decision artifact is a versioned license register: an inventory that connects each model, dataset, and software component to its provenance, governing terms, intended use, obligations, and accountable reviewer. Keep it in your release repository beside the build definition; no NVIDIA product generates it. It makes the release decision reproducible when a model, dependency, or delivery channel changes.
How it works
Identify the actual artifact and transaction
Record the exact model revision, dataset snapshot, and software release or commit, together with checksums where your release process supplies them. Preserve the retrieved license text and its date alongside the source URL. A family name or repository homepage is insufficient when several artifacts travel together.
For each entry, describe what leaves your organization (model files, modified code, dataset rows, generated answers, or hosted API access), to whom, and for what commercial purpose. The licensing owner can then assess the relevant grant instead of the vague question, “Can we use this model?”
NVIDIA-Nemotron-Nano-9B-v2 illustrates the separation. Its model card links the NVIDIA Open Model License Agreement for model use and separate NVIDIA API Trial Terms of Service for the trial endpoint. Approval of downloaded weights does not approve the hosted trial transaction.
A self-hosted NIM adds a third instrument. The NIM for LLMs legal page states, “By using this NIM, you acknowledge that you have read and agreed to the NVIDIA AI Product Agreement.” It also states that materials the software retrieves “are not distributed with this software and are governed solely by separate terms, conditions and licenses.” The inference container and the model weights it pulls are therefore two register entries with two reviews.
Cards differ in how many licenses they name. The Llama 3.1 Nemotron Safety Guard 8B v3 card governs model use by the NVIDIA Open Model License Agreement, then adds “Additional Information: Llama 3.1 Community License Agreement” and “Built with Llama”. That upstream agreement and its attribution join the review, so read the exact terms each chosen card links instead of carrying a familiar label over from another model.
Read the model’s grant and conditions
The NVIDIA Open Model License Agreement grants use, modification, sale, and distribution subject to its conditions. Section 3.1 requires that recipients of a distributed Model receive a copy of the agreement and a Notice text file containing the attribution “Licensed by NVIDIA Corporation under the NVIDIA Open Model License”. Section 4 addresses separately licensed components; keep their notices in the review. The agreement claims no NVIDIA ownership of outputs, while assigning responsibility for outputs and their subsequent uses to you.
Two conditions limit the use itself. Under Section 2.1, if you bypass, disable, or reduce the efficacy of a Guardrail contained in the Model “without a substantially similar Guardrail appropriate for your use case, your rights under this Agreement will automatically terminate.” Under Section 2.3, use “must be consistent with NVIDIA’s Trustworthy AI terms”. A fine-tuning or serving plan that could weaken a safety mechanism therefore goes to the licensing owner before the work starts, together with the replacement Guardrail it proposes.
Translate the approved interpretation into release checks: the files to ship, notices to retain, permitted operations, and changes that require another review. The release engineer owns packaging; the licensing owner resolves ambiguous terms.
Keep dataset rights separate
NVIDIA NeMo Gym’s math_with_judge README labels its code Apache 2.0, OpenMathReasoning data Creative Commons Attribution 4.0 International, and Math Stack Overflow data Creative Commons Attribution-ShareAlike 4.0 International. That single directory demonstrates why the code license cannot serve as the dataset license.
For material actually licensed under CC BY 4.0, Section 3(a) specifies attribution when sharing: preserve supplied creator and notice information, identify the license, retain a practicable source link, and indicate modifications. Section 2(b) distinguishes privacy and publicity rights from the copyright permission. A dataset’s reuse license therefore does not settle whether personal information may be processed for your purpose.
Track source datasets separately even after filtering or concatenation. If a proposed release contains ShareAlike material, send that material and its intended adaptation and sharing operation for review under its own terms. Do not silently relabel the mixture CC BY because another input uses that license, and read the ShareAlike legal code itself rather than reusing the CC BY checklist above.
Preserve software redistribution obligations
NeMo Agent Toolkit v1.9.0 ships an Apache License 2.0 in LICENSE.md. Section 4 requires a license copy for recipients, prominent notices in modified files, retention of applicable source notices, and preservation of relevant attribution from a supplied NOTICE file under its stated conditions. Those requirements become packaging acceptance criteria for a distributed derivative.
Keep the application’s dependency inventory beside that record: approving the toolkit’s license does not cover every package in your appliance, so inspect each component’s own terms in the same release review. Check these obligations in the packaging pipeline, against the files in the image; no runtime setting satisfies them.
Key distinctions and trade-offs
Use the artifact and operation together to decide what to inspect:
| Artifact | Deciding condition | Mechanism to inspect | Delivery cost |
|---|---|---|---|
| Model weights | Internal inference, modification, or distribution | Model card’s governing agreement and separate-component notices | Review and packaging work per approved model revision |
| Inference container | Running or shipping a self-hosted NIM | NVIDIA AI Product Agreement named on the NIM legal page | Separate review from the weights the container pulls |
| Dataset | Training, evaluation, adaptation, or sharing | Dataset-specific license and source attribution | Provenance and notice maintenance per input |
| Software | Use of original code or redistribution of modified code | LICENSE.md, applicable NOTICE, and dependency terms | Dependency inventory and release-file checks |
| Hosted model access | Requests go to a provider’s service | Service terms linked for that endpoint | Separate service approval and operational commitments |
An internal service and a customer appliance raise different distribution questions, and each independently licensed component adds review work even when it is permissive.
Architecture / decision scenario
An industrial-equipment company builds a maintenance assistant for 40 customer sites. Its initial plan is a hosted service; procurement then requests 12 disconnected appliances, with ten working days before the first shipment. Because the appliances are air-gapped, the weights must be pre-staged in the image, so the company ships them itself. The bundle holds a NIM for LLMs container, NVIDIA-Nemotron-Nano-9B-v2 weights, modified NeMo Agent Toolkit code, and an on-site self-test set that mixes 300 of the company’s own resolved tickets with 500 rows from a public CC BY 4.0 troubleshooting corpus.
The tempting shortcut is to reuse the hosted-service approval and put one Apache license in the image. The release engineer rejects it: the image carries four separately licensed artifacts. The licensing owner budgets two working days per third-party instrument, so the AI Product Agreement, the Open Model License, the toolkit’s Apache license, and the corpus’s CC BY terms take eight of the ten days and leave two for packaging and notice checks. The safety team then asks to add Llama 3.1 Nemotron Safety Guard 8B v3 as a content-safety check. Its Open Model License terms are already under review, but the Llama 3.1 agreement it names is a fifth instrument, and twelve days of work no longer fit in ten.
The product owner decides scope. Replacing the 500 public rows with 500 more of the company’s own resolved tickets removes the corpus review and returns two days, enough for the Llama 3.1 agreement; deleting the corpus attribution while keeping its rows is not an option. The alternative is to ship without the Safety Guard model and add it in the second shipment. Shipment proceeds only after the register matches the image contents; unresolved rights go back to the licensing owner before release.
Common exam traps
- Treating downloadable weights as unconditional commercial permission.
- Applying a repository’s software license to datasets in that repository.
- Assuming a model-use grant also approves a hosted trial service or the NIM container.
- Bypassing a model’s Guardrail without a substantially similar replacement, which terminates the Open Model License rights.
- Removing required notices because an appliance runs offline.
- Treating output ownership wording as clearance of third-party rights.
- Copying a dataset-sharing checklist to a different license without review.
What to remember
Assess each artifact against its actual use and distribution path. Preserve the exact terms, provenance, notices, reviewer, and approved operations in the release record. Model, dataset, software, and service permissions remain separate even when one application packages them together.
Knowledge check
Your license register already approves the NVIDIA-Nemotron-Nano-9B-v2 weights under the NVIDIA Open Model License for internal inference. The team now plans to serve those same weights from a self-hosted NIM for LLMs container on its own cluster. What must the register gain before that deployment goes live?
A team plans to fine-tune a model governed by the NVIDIA Open Model License. Its own evaluation shows the fine-tune would reduce the efficacy of a safety Guardrail contained in the model, because that Guardrail refuses some legitimate high-voltage maintenance questions. What should happen before the fine-tuning starts?
A commercial product will ship an evaluation set built from the NeMo Gym math_with_judge resources. That directory labels its code Apache 2.0, its OpenMathReasoning data CC BY 4.0, and its Math Stack Overflow data CC BY-SA 4.0. The set keeps rows from both datasets after filtering. How should the license register record it?
An air-gapped appliance distributes NVIDIA-Nemotron-Nano-9B-v2 weights governed by the NVIDIA Open Model License and modified NeMo Agent Toolkit source code under the Apache License 2.0, including applicable source notices and a supplied NOTICE file. Because the appliance never reaches the internet, a packaging engineer proposes trimming its license material. Which packaging check meets both licenses?
Notice , modified-file notices, and applicable code attributionsNotice file with NVIDIA’s attribution, and Apache 2.0 Section 4 requires a copy of the license, prominent notices in modified files, applicable source notices, and relevant attributions from the supplied NOTICE file. Being offline does not remove those obligations, a URL is not a copy, the toolkit’s license does not cover the weights, and these are checks on the files in the image that no runtime setting satisfies.A self-hosted appliance already has an approved Open Model License review for its main model. The safety team adds Llama 3.1 Nemotron Safety Guard 8B v3, whose model card governs model use by the same NVIDIA Open Model License Agreement and lists the Llama 3.1 Community License Agreement under “Additional Information”, with “Built with Llama”. What does adding it to the release require?
References
- NVIDIA, NVIDIA-Nemotron-Nano-9B-v2 model card
- NVIDIA, NVIDIA Open Model License Agreement
- NVIDIA, NeMo Agent Toolkit v1.9.0 LICENSE.md
- NVIDIA, NeMo Gym math_with_judge licensing information
- NVIDIA, Legal and Acknowledgements (NVIDIA NIM for LLM and VLM)
- NVIDIA, Llama 3.1 Nemotron Safety Guard 8B v3 model card
- Creative Commons, Attribution 4.0 International legal code