Link to this sectionOn Premise#

On Premise connects CPU and optional NVIDIA GPU workers on your own Linux, Apple Silicon macOS, or Windows host to Ultralytics Platform. Platform remains the hosted control plane for the UI, authentication, metadata, annotations, and job orchestration, while every pixel and trained model artifact stays on your premises.

Your host needs Docker and outbound HTTPS access to Platform. The installer adds Docker automatically when it is missing, so the normal setup is one command.

Link to this sectionSystem Requirements#

CPU ingest and training work on all three operating systems. The installer selects the official native arm64 image on Apple Silicon and ARM Linux, so small jobs such as YOLO26n on COCO8 run without x86 emulation. NVIDIA acceleration is optional; when it is unavailable, training runs on CPU.

Link to this sectionData Boundary#

Dataset folders are mounted read-only. Platform and its hosted workers never receive the source or derived pixels, and On Premise jobs never fall back to Ultralytics or RunPod compute.

Link to this sectionConnect a Host#

1. Open Ultralytics Platform on the Linux, Apple Silicon macOS, or Windows host that can access your datasets.
2. Go to Settings > Integrations and select Connect on the On Premise card.
3. Platform selects the detected Linux, macOS, or Windows command. Apple Silicon is required on macOS. Keep the prefilled values or change them:
   • Machine name: On Premise host
   • Dataset folder: /datasets on Linux or ~/Ultralytics/datasets on macOS and Windows
   • Models folder: /models on Linux or ~/Ultralytics/models on macOS and Windows
4. Select Create install command. The dialog tells you which terminal to open for the selected operating system.
5. Copy the complete command, paste it into that terminal, and run it. The command includes the one-time enrollment token, installs and starts Docker when needed, and creates the selected folders.
6. Leave the dialog open. Platform checks every 500 milliseconds and shows the host as connected when the CPU worker starts. A GPU worker starts automatically when Docker exposes a supported NVIDIA runtime.

The enrollment token expires after 10 minutes and can be exchanged only once. The installed worker stores the resulting revocable worker key in a mode-0600 environment file. It never receives Platform MongoDB or cloud-storage credentials. Compose restarts the workers automatically, and setup configures Docker to start at boot on Linux or sign-in on macOS and Windows.

Link to this sectionCreate an On Premise Dataset#

1. Put the dataset beneath the connected dataset folder. For example, /datasets/warehouse is warehouse inside the default root.
2. In Platform, select New Dataset > On Premise.
3. Browse the connected host with the same folder browser used for Google Cloud Storage, Amazon S3, and Azure Blob Storage, select a folder, choose the task, and create the private dataset.
4. The host indexes the dataset and reports metadata. Platform never uploads the images.

On Premise uses the same CPU ingest code as hosted uploads. It supports:

• loose images and videos;
• ZIP, TAR, TAR.GZ, and TGZ archives;
• Ultralytics NDJSON and COCO JSON;
• YOLO datasets and classification folder layouts; and
• detect, segment, pose, OBB, and classify tasks, including the same class mapping, task inference, validation, and split handling.

The storage output is the only difference. Hosted ingestion may resize or normalize images and create thumbnails in Platform storage. On Premise never resizes, re-encodes, edits, or deletes mounted originals. Archive contents, remote NDJSON assets, and video frames sampled at 1 FPS up to 100 frames, then evenly across longer videos, are written only to a Docker volume on the host.

Link to this sectionPreview and Annotate#

Platform authorizes each preview, then your browser loads the revision-bound file directly from http://localhost:8765 on the same computer. No hostname, certificate, VPN, proxy, or preview setting is required.

Annotations are stored as Platform metadata. Editing or deleting an image in Platform changes the Platform reference and annotations only; it never changes a source file or label sidecar.

Link to this sectionTrain Locally#

Start training from the normal project training dialog. A dataset bound to an On Premise host is claimable only by that host. Platform uses its GPU worker when available and otherwise runs the same training code on its CPU worker. Training reads the mounted files, writes checkpoints and weights beneath the configured models folder, and returns job state, scalar metrics, and the immutable checkpoint reference to Platform. Model downloads use the same signed localhost connection as previews, so the weights move directly from your host to your browser.

On Premise training does not consume Platform compute credits. Ultralytics hosted workers and RunPod cannot claim the job or read its pixels or artifacts.

Link to this sectionManage the Worker#

Use the On Premise card in Settings > Integrations to view CPU/GPU availability, reconnect a host, or disconnect it. Reconnecting rotates the worker secret without changing existing dataset identity. Disconnecting revokes future claims and preview access; it does not delete datasets, source files, cached pixels, or model artifacts from the host.

To inspect or stop the installation on Linux:

cd /opt/ultralytics-worker
docker compose logs -f
docker compose down

On macOS and Windows, the installer prints the equivalent command using ~/.ultralytics/worker.

Also see Datasets, Annotation, and Cloud Training.