Model Settings

You can turn on additional controls for your workspace on how models use your data when you run your agents. These limit the model choices available to you.

​

No training

Only use providers whose agreements or API terms prohibit training on your data. Providers may still retain data for other purposes, such as abuse monitoring. Skydive will never train its own models to improve agents on your data.

​

Zero data retention (ZDR)

Only use providers configured to retain none of your prompts or responses, including for abuse monitoring. This is the strictest setting and offers the fewest models.

​

Google integrations

Agents connected to Google services automatically use models that meet Google’s Limited Use requirements, including no training on your Google data. As a result, model choice is limited further.

​

Which models can I use?

Skydive offers 300+ models to run your agents, including frontier models from OpenAI, Anthropic, xAI, and Google and leading open-source models. You can train your agent on any model and it retains its memories when you switch. Over time, you can lower costs by moving it to an open-source model. In some cases, the model costs one-tenth as much. Skydive Glide is our default model router. It chooses the best model for your task and balances cost and quality without you needing to switch. Controls like No Training and ZDR apply at the routing layer. Each model can be served by multiple providers with different data terms, so availability depends on the provider, not just the model’s name. Open-source models, including those behind Glide, can qualify. These controls apply to every request and fallback. If no eligible provider is available, the request fails rather than using one that violates your settings. Check your agent’s model picker for the models available to you. These controls govern model-provider handling of your data, not your conversation history in Skydive.