Skip to main content

​
What Together stores

By default, Together stores the prompts you send and the responses models return, and may use them for product improvements. This data is not shared with third parties, and you can delete it at any time. Organization admins can turn storage off in privacy settings, which enables zero data retention (ZDR): request content is then not persisted or used for any secondary purpose. See zero data retention for what ZDR covers, how to enable it, and what Together retains.

​
Training opt-in

Data sharing for training other models is opt-in and not enabled by default. Check or change this setting in the Privacy section of Organization Settings. See the privacy policy for the full legal picture.

​
Organization privacy settings

Only organization admins can change the three toggles in the Privacy section of Organization Settings:
  • Prompt storage: On by default. Stores prompts and model responses for product improvements. Turn it off to enable ZDR, which also turns off passthrough models.
  • Passthrough models: On by default. Allows your organization to use models that forward prompts and responses to third-party providers. Passthrough requires prompt storage, so you can’t turn this toggle on while prompt storage is off.
  • Model training: Off by default. Allows your organization’s data to be used for training models released by Together AI and partners.
If your organization is on the Limited tier, add a payment method before you change these settings.

​
Account vs. organization settings

You may see a privacy toggle on both your personal account profile and your organization settings. These control different scopes:
  • The account setting applies only to traffic you send under your personal account when it isn’t attached to an organization.
  • The organization setting governs all traffic sent under that organization’s projects and API keys, regardless of which member makes the request.
When a request uses an organization’s API key, the organization setting is what applies. To turn data sharing off for your team, change it in organization settings, not on your personal profile.

​
Passthrough third-party models

Some models are offered as passthrough, which means Together forwards your prompts and responses directly to the upstream provider. The provider handles that data under its own data policy. The Passthrough models toggle in organization privacy settings controls whether your organization can use these models, independent of the training opt-in. To keep all traffic on Together’s infrastructure, turn it off. Non-passthrough models keep working as normal.

​
Enterprise data residency and private networking

For select serverless models, geo-specific inference keeps inference processing in the United States, with no GPUs to provision. The guarantee applies to requests sent to us-api.together.ai. It covers only where those requests are processed, not where data is stored, logged, or routed. For other data-residency, regulatory, or compliance requirements, such as GDPR, Together supports private networking and virtual private cloud (VPC) deployments, including in EU regions. Use a dedicated endpoint, or contact sales to discuss the right setup for your workload.

​
Vulnerability disclosure

Together AI accepts security vulnerability reports through its HackerOne program. The program is private, so viewing it and submitting reports requires a HackerOne account. Program details are published in the standard security.txt file. If you believe you’ve found a security issue, submit it through HackerOne rather than a support ticket.

​
Third-party model providers

Models published by third-party authors (DeepSeek, Qwen, Mistral, etc.) and hosted on Together run on Together’s own infrastructure. They do not call out to the model author. The model author has no access to your requests or API calls. For example, DeepSeek models are hosted in Together’s secure North America data centers. DeepSeek itself receives no user requests or API traffic from this deployment. Models on Together are hosted at full precision. Together does not distill them, force system prompts, or layer censorship on top. The version you call is the version the model author published.