“Free” is a licensing claim, not a price

What AGPL-3.0 section 13 actually says

Verbatim, from https://www.gnu.org/licenses/agpl-3.0.txt.

13. Remote Network Interaction; Use with the GNU General Public License. Notwithstanding any other provision of this License, if you modify the Program, your modified version must prominently offer all users interacting with it remotely through a computer network (if your version supports such interaction) an opportunity to receive the Corresponding Source of your version by providing access to the Corresponding Source from a network server at no charge, through some standard or customary means of facilitating copying of software.

Three conditions must all hold. They are worth taking one at a time, because the engineer in the room who challenges this is challenging the first one.

“Mere interaction is not conveying” — section 0, and why it does not save you

To “propagate” a work means to do anything with it that, without permission, would make you directly or secondarily liable for infringement under applicable copyright law, except executing it on a computer or modifying a private copy. Propagation includes copying, distribution (with or without modification), making available to the public, and in some countries other activities as well. To “convey” a work means any kind of propagation that enables other parties to make or receive copies. Mere interaction with a user through a computer network, with no transfer of a copy, is not conveying.

Under GPL-3.0, running a modified copy on your own server and letting people call it over HTTP is not conveying, so no source obligation arises. That is the ASP loophole. AGPL-3.0 section 13 exists specifically to close it: the duty is triggered by network interaction rather than by distribution. An internal, closed-source inspection microservice, modified in any way and exposed on the plant network, is the exact scenario the clause was drafted for. The obligation is to offer source to those users, not to the world, but sections 2 and 10 then let any of those users redistribute it freely, so the practical effect is public release.

Test your own deployment

A faithful walk through the three conditions. A reading of the licence text, not legal advice.

What the vendor itself says, which is broader than the licence

• ultralytics.com/license — An Enterprise License is required to use the model without open-sourcing your entire project. Internal business tools, private company applications, SaaS platforms, embedded hardware and edge devices, and custom-trained models in commercial settings are all enumerated as Enterprise triggers.
• docs.ultralytics.com AGPL compliance section — Compliance requires publishing the complete larger application, any modifications, training and inference scripts, configuration, backend and frontend code, modified third-party libraries, training data if needed to retrain, and model weights if modified or fine-tuned. Note the weights clause: the vendor treats a fine-tuned defect model as part of the corresponding source, so your own inspection weights would have to be published.
• ultralytics/ultralytics issue #19390, February 2025 — A maintainer, on an explicitly non-network on-device product: you would very likely need the enterprise license if you are not intending to release the source for the app that is integrating the model.
• ultralytics/ultralytics issue #22458 — The gap between the licence text and the vendor's reading is contested in public, under the title “License Ambiguity: Ultralytics Website does not follow the spirit of GNU AGPLv3”.

The enterprise licence, on its published terms

Detection models: licence, obligation, accuracy and latency

Rows marked SAB were measured by Roboflow's single-artifact benchmarking harness on the full 5,000-image COCO val2017 split with pycocotools, NVIDIA T4, TensorRT, FP16, batch 1, so they are mutually comparable. Rows marked “authors” are vendor-reported and are not directly comparable — they are labelled rather than quietly mixed in.

Two traps a licence badge will not catch

And the same lesson in the data

The most thematically perfect dataset in the entire search — submersible pump impellers, which are as close to a compressor component as public data gets — is CC BY-NC-ND 4.0. NoDerivatives is fatal: you may not publish even the re-labelled derivative this demonstration is built on. The most relevant dataset is the least usable, and that is the single most memorable fact in the data audit.