Zero data retention

Qwen Cloud does not use your API inputs or outputs to train or improve models. This page explains how your data is handled during the inference process.

​

Default data handling

When you send an API request to Qwen Cloud:

• Inputs and outputs are processed in memory for the duration of the request. Once the response is returned, the data is not retained in persistent storage.
• Metadata such as token counts, timestamps, and request IDs is logged for billing, rate limiting, and service delivery.
• No training on your data: Your prompts and completions are never used to train, fine-tune, or improve Qwen models.

​

Prompt caching

When context caching is active, cached prompt data and associated key-value caches remain in volatile memory for the duration of the cache TTL. Cached data is not written to persistent storage and is automatically evicted when the TTL expires.

​

Conversation storage (Responses API)

The Responses API stores conversation context for multi-turn workflows:

• Context linking: Use previous_response_id to link turns in a conversation. The response ID remains valid for 7 days, allowing the server to automatically retrieve and combine previous inputs and outputs as context.
• Automatic expiration: Conversation data linked through previous_response_id is automatically deleted after 7 days.

This storage is limited to the Responses API. All other Qwen Cloud APIs follow the default zero-retention behavior described above.

​

Best practices

• Be aware that conversation context is stored for 7 days when using previous_response_id in the Responses API. Avoid linking turns if your application handles highly sensitive data.
• Avoid including personally identifiable information (PII) in prompts when possible.
• Implement data minimization at the application layer — send only the context the model needs.

​

Learn more

• Data security: Encryption, API key security, and content moderation.