Does the Ultralytics Python Package Transmit Any Data From Local Devices by Default?

[Brayan532]

Hi Ultralytics team,

I’d like to clarify what information, if any, the ultralytics Python package automatically sends from a user’s local machine during ordinary usage, without the user explicitly connecting to services such as Ultralytics HUB, W&B, ClearML, or Comet.

My understanding is that standard Ultralytics executions may send anonymous usage telemetry such as:
▪ Ultralytics/Python version
▪ task type
▪ CPU/environment information
▪ device type
▪ whether the model is custom
▪ a random/session identifier
Could you confirm whether this is accurate and provide the complete set of information that may leave the local device by default?
In particular, does the package ever automatically transmit any of the following:
▪ file names or file paths
▪ model names or model weights
▪ images, videos, datasets, or inference inputs/outputs
▪ usernames, hostname, IP-related information, or other device identifiers
▪ environment variables
▪ project/directory names
▪ command-line arguments
▪ source code or configuration files
▪ any other information that could identify the user, machine, project, or data

Also, could you clarify:
a. Which endpoints/domains receive the default telemetry?
b. Whether telemetry is enabled by default.
c. How it can be completely disabled.
d. Whether disabling telemetry guarantees that no network requests are made by the Ultralytics package itself, assuming integrations such as HUB, W&B, ClearML, and Comet are not being used.
e. Where in the source code the telemetry/network behavior is implemented so that users can independently audit it.

I'm particularly interested in understanding whether the package can transmit anything from a personal/local device in the background beyond the documented anonymous telemetry.

Best wishes,
Brayan.

[UltralyticsAssistant]

👋 Hello @Brayan532, thank you for the detailed privacy and telemetry questions. You’re asking for a complete, version-specific accounting of default network requests, transmitted fields, endpoints, disabling behavior, and the relevant source-code locations; an Ultralytics engineer will verify these details and assist soon. This is an automated response in the meantime.

[glenn-jocher]

Yes—sync=True by default. Current analytics POST to Google Analytics includes a hashed machine-derived UUID, random session ID, package/Python/PyTorch/environment/CPU information, task/mode, device, model basename, and mode-specific fields such as dataset stem, architecture, GPU, image count/size, timings, metrics, optimizer, batch, epochs, and export format; it does not send weights, media, inference outputs, full paths, source/config contents, environment variables, hostname, username, or full CLI arguments. Sentry may additionally receive crash/error context only when sentry-sdk is already installed and its other guards pass. Disable both with yolo settings sync=False; however, this does not guarantee zero networking because import performs DNS-based connectivity checks and requested or missing models/assets may be downloaded. For offline operation also set YOLO_OFFLINE=1, avoid remote resources/integrations, and enforce network isolation if a strict guarantee is required. See the privacy documentation, analytics implementation, and settings, online-check, and Sentry implementation.

[Brayan532]

Dear Glenn,
Thank you for your clarification.
I'm wondering what else has not been sent out.
Particular, could you clarify what do you mean by model basename,

And architecture, metrics, epochs in your statement as
.....and mode-specific fields such as dataset stem, architecture, GPU, image count/size, timings, metrics, optimizer, batch, epochs.

Could you clarify whether you are monitoring the basename, it means the name we have given to our model?
And what about architecture? What does it mean? Or metrics?
So, any results we obtain in our device, you are monitoring and aware of that as immediate as we are.

Best wishes,
Brayan.

[glenn-jocher]

model basename is only the final filename, e.g. /private/project/my_model.pt becomes my_model.pt, so a custom filename is transmitted but not its path. architecture is the inherited model family/config identifier, e.g. yolo11n-seg; epochs_done is the number completed in that training session; and the training fitness field is one rounded task-specific summary score, not all losses, per-class metrics, predictions, or result files. Events are sent asynchronously after supported operations and analyzed as anonymous aggregate usage telemetry—not as live access to your images, weights, predictions, or complete local results. Set yolo settings sync=False to disable these analytics.