Ultralytics YOLO licence for companies: AGPL-3.0, Enterprise or Apache 2.0 detectors

The licence question in a production setting

A quality or production team that wants YOLO for visual inspection inside the company has two paths with Ultralytics: the free AGPL-3.0 licence, which Ultralytics describes as "the right one for you if you are comfortable open-sourcing your entire project under AGPL-3.0", or a paid Enterprise Licence. Ultralytics' licence page names internal business tools and custom-trained models in a proprietary setting among the uses that require the Enterprise Licence. A company that wants neither has detectors under the permissive Apache 2.0 licence as a third path.

This article is general information, as of 25 October 2026, not legal advice. It quotes what Ultralytics' licence page and the AGPL-3.0 text say, and your legal counsel confirms how they apply to your project. For the licences of language models, see our separate guide to open LLM licences for company use.

A short explanation of the terms for readers outside IT. YOLO ("You Only Look Once") is a family of object detection models: they find and outline objects such as parts, defects or people in an image. Ultralytics is the company that develops the YOLOv5, YOLOv8, YOLO11 and YOLO26 versions and publishes the software that trains and runs them. AGPL-3.0, the GNU Affero General Public License version 3, is an open-source licence with a "copyleft" condition: software derived from it has to be passed on under the same licence, with its source code. Apache 2.0 is a permissive licence: it grants the right "to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work", and asks whoever redistributes it to pass on the licence text, mark changed files and keep the copyright and attribution notices.

What Ultralytics' licence page states

Ultralytics publishes its position on one licence page. The core sentence is: "An Enterprise License is required if you want to use Ultralytics YOLO without open-sourcing your entire project." The page then lists the uses for which it requires that licence.

Ultralytics licence page, ultralytics.com/license, read on 25 October 2026 (no date on the page; © 2026). Wording shortened: "Internal business tools or private company applications", "Embedded deployments in hardware, edge devices, robotics, cameras, or appliances", "Using custom-trained or fine-tuned YOLO models in a proprietary or commercial setting", "R&D projects that are not fully open-sourced", "Customer-facing solutions where you do not want to publish code, configs, training scripts, and related components".

On models, the page states: "All Ultralytics YOLO trained models fall under the AGPL-3.0 License by default" and "The AGPL-3.0 License covers the training code and the models produced by that training code." Its FAQ adds that AGPL compliance can require releasing model weights where applicable, as well as the larger application, modifications, scripts and configuration files, and that this applies even when a model is trained from scratch, used only internally or for R&D, or deployed through an API.

On the Enterprise side, the FAQ describes the licence as giving organisations "the ability to modify Ultralytics YOLO source code and embed Ultralytics YOLO code and models in commercial products". It covers "the complete Ultralytics YOLO source code portfolio, including YOLO26, earlier YOLO versions, and any future YOLO models that Ultralytics may release during the license term", can be scoped to one project or the whole organisation, and is priced per organisation and use case through a request form. The GitHub repository of the ultralytics package shows the AGPL-3.0 licence file and the same two options in its README.

What the AGPL-3.0 text itself says

The licence page is Ultralytics' description of how it applies its licence. The legal text is the GNU Affero General Public License, "Version 3, 19 November 2007". Three of its provisions are the ones companies ask about.

Section 0 defines conveying: "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." Section 2 states: "You may make, run and propagate covered works that you do not convey, without conditions so long as your license otherwise remains in force." Section 13, the clause that distinguishes the Affero licence from the ordinary GPL, states that "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".

The two documents do not read the same way at first sight, and we do not reconcile them here. Ultralytics is the licensor and states which uses it licenses commercially; the AGPL text sets the conditions of the open-source licence. For a concrete project, counsel looks at both, together with facts only you know: whether code or models leave the company (to a camera supplier, a customer, a subsidiary), whether the Ultralytics code was modified, who interacts with the system over a network, and how the trained weights are used.

Whatever the outcome, the hardware stays the same, and we supply it for every path; tell us how many lines and cameras the inspection covers, and our team replies with the cards and server that fit.

Code, weights and training data are licensed separately

A detection model in production is three things: the code that trains and runs it, the weights it learned, and the images it learned from. Each can carry its own licence, and the open-source detectors show it.

The D-FINE repository is under Apache 2.0, but its README notes that checkpoints trained or pretrained on Objects365 "may be subject to the Objects365 dataset terms" and "should not be assumed to be commercially cleared under the D-FINE license". The RF-DETR repository states that "The open-source rfdetr package and Apache-designated model weights are licensed under Apache License 2.0", while its "Plus components, including the rfdetr_plus extension and RF-DETR-XL / RF-DETR-2XL detection models, are licensed under PML 1.0"; its model table also marks the A, F and P sizes as PML 1.0. COCO, the dataset most pretrained detectors start from, is marked CC BY 4.0 in Ultralytics' dataset documentation.

Retraining adds one more point. Each retrained model is a new set of weights produced by the same training code, and Ultralytics' statement that its licence covers "the models produced by that training code" applies to every such version, not only to the first one that went to the line. It helps to take the licence decision once per toolchain and record it with each model version.

For inspection projects the training images are usually the company's own photos of its own parts, which removes one licence question. They can show employees, which makes them a data protection matter for your data protection officer.

Detectors under Apache 2.0

GitHub repositories lyuwenyu/RT-DETR, Peterande/D-FINE, roboflow/rf-detr and Megvii-BaseDetection/YOLOX, licence shown in each repository and README notes, read on 25 October 2026. "In Transformers": the RT-DETR README states "RTDETR and RTDETRv2 are now available in Hugging Face Transformers".

Moving from Ultralytics to one of these projects means a new training pipeline, new hyperparameters and a new validation run against your inspection criteria, since accuracy figures from COCO do not transfer to a production line. A team that already runs Ultralytics in a pilot can weigh the cost of that work against the Enterprise Licence; a team that starts from scratch can choose the licence first and the model second.

Steps before a YOLO model goes to the line

1. List where Ultralytics code runs: training server, inspection PCs, cameras, cloud services.
2. Note who receives copies of code or weights outside your legal entity, such as a machine builder, a camera supplier or a customer.
3. Record whether the Ultralytics code was modified and whether users reach the system over a network.
4. Have legal counsel review these facts against Ultralytics' licence page and the AGPL-3.0 text.
5. Decide: Enterprise Licence from Ultralytics, publication under AGPL-3.0, or a detector under Apache 2.0, with the weight licences checked per checkpoint.
6. Keep the licence record with the model version, so a retrained model inherits the same decision.

The GPU plan is the same on every licence path

The licence decides which software you run, not which hardware it needs. Detection models are small: Ultralytics lists 55.7 million parameters for YOLO26x and 2.4 million for YOLO26n, and D-FINE's largest size has 62 million. GPU memory goes to batch size and image resolution during training, and to batches and model copies during serving.

For inference next to the line, one NVIDIA RTX PRO 4500 Blackwell Server Edition, a passive single-slot card with 32 GB at 165 W, or an L4 at 72 W in a low-profile slot covers most single-site projects. In NVIDIA's DeepStream 9.2 figures, RT-DETR at 640 × 640 in FP16 ran at 380 frames per second on an RTX PRO 4500 Server Edition and 1,037 on an RTX PRO 6000 Server Edition. Training uses the same cards, or a server with two or four RTX PRO 6000 Server Edition cards for many experiments, as our guide to the GPU server for training vision models works out. The serving stack, from ONNX export to TensorRT and Triton, is in our guide to serving vision models.

We supply the cards and servers for both paths, but no software licences from Ultralytics and no licence advice; send us your camera count, image sizes and the models you plan, and our team replies with a hardware proposal.

Which card for which inspection setup

NVIDIA product pages and MIG guide as summarised in our research notes of 24 October 2026; DeepStream 9.2 performance page, read on 25 October 2026. Card choice is our recommendation.

Before a card goes into a server you already run, check slot power, airflow and the maker's option list; our checklist for adding a GPU to an existing rack server covers these points. Inspection and anomaly detection without in-house data scientists is the subject of our guide to visual inspection AI.

What we supply

We supply the NVIDIA L4, the RTX PRO 4500 Server Edition and the RTX PRO 6000 Server Edition as cards for your servers, and AI servers built to order, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. On request the server arrives with the operating system, NVIDIA driver, CUDA and a container runtime. Quantities and lead time are confirmed in writing with the quote.

FAQ