Data retention

Inference.net is built to support production workloads without treating request data casually.

#Core principles

• Request data is not used for model training by default.
• Secrets and similar sensitive values are stripped where possible.
• Platform data is encrypted in transit and at rest.
• Retention should match the operational need of the workflow.

#Zero Data Retention

New teams start with the Zero Data Retention (ZDR) setting on, so Inference.net does not retain request and response content. With ZDR on, message contents, tool and function schemas, custom metadata, and error details are purged after processing; model, usage, and timing metadata are retained.

Change the setting in team settings in the dashboard. Turn it off to keep inference results or to use the Asynchronous and Batch APIs. See the Zero Data Retention Policy for what ZDR covers.

#Retention and asynchronous inference

The Asynchronous and Batch APIs store requests and results until they are delivered, so they are unavailable while ZDR is on. Teams with ZDR on receive a 403 Forbidden on /v1/slow, /v1/batches, and /v1/files upload requests, and should use the synchronous /v1 endpoints or turn ZDR off in team settings.

#Ad-supported models

Ad-supported models (for example moonshotai/kimi-k3.1-pro-with-ads) can include sponsored content from our advertising partner, Gravity. When you call an ad-supported model, Inference.net sends Gravity, for a sample of requests:

• the full request: chat messages (images included) and parameters such as tools and temperature;
• the full response the model generated, including any reasoning and tool calls;
• a SHA-256 hash of the user's email address (never the address itself);
• the IP address the request came from;
• the end-user id you send in the request's safety_identifier or user field, if any;
• the conversation id you send in the x-session-id header or the request's session_id field, if any;
• the request's token usage, finish reason and cost (the amount billed, the list price, and the difference).

This happens whether or not ZDR is on for your team: it is a term of using an ad-supported model. To keep all request content under ZDR, call the standard model instead. Ad-supported models also have spend limits.

#Need a specific retention policy?

If you need a specific policy, a data processing agreement, or help mapping Inference.net into your internal compliance requirements, contact our team.