Is it allowed to distill Deepseek on https://ai.azure.com/?

DeepSeek can be deployed for inference in Microsoft Foundry, but the available Microsoft documentation does not establish blanket permission—or built-in support—to distill every Azure-hosted DeepSeek model.

Check the exact DeepSeek deployment before proceeding

1. In Microsoft Foundry → Discover → Models, open the exact DeepSeek model and version you intend to use.
2. Open its License tab and review the Pricing and terms presented during deployment. Foundry classifies non-Microsoft models as non-Microsoft products subject to the terms supplied with each model; customers must review the model card, license, and related documentation themselves.^1^
3. Look specifically for terms governing:
   • reuse of generated output as training data;
   • training or improving another model;
   • distillation or creation of derivative models;
   • commercial use and redistribution.
4. Check the Foundry fine-tuning supported-models table for both models involved. Support for inference, fine-tuning, or one training method does not automatically imply support for distillation. Foundry documents distillation through teacher-generated supervised fine-tuning data, but says availability is constrained by the model, approach, method, and modality.^2^
5. If either the license or supported-model matrix does not expressly cover the intended operation, obtain confirmation through Microsoft Support before creating the dataset—particularly for commercial use.

Do not assume that different DeepSeek variants have identical licenses. The model and version selected in Foundry determine which model card and terms apply.

Built-in distillation versus custom distillation

The legacy Stored Completions → Distill workflow applies only to Azure OpenAI stored completions in Foundry (classic). It requires at least 10 stored completions, creates an internal JSONL training dataset, and does not permit that generated file to be downloaded. The feature is deprecated and scheduled to retire on October 15, 2026.^3^

That documentation does not identify DeepSeek as a supported teacher for this workflow. An earlier Microsoft Q&A answer likewise reported that Foundry’s model-distillation feature was limited to selected models and that no DeepSeek expansion had been announced at that time.^4^

Consequently:

• DeepSeek available for deployment/inference: Yes. Microsoft documents deploying DeepSeek and sending inference requests.^5^
• DeepSeek available as a teacher in every built-in distillation workflow: Not documented.
• Using DeepSeek outputs to build a dataset outside that workflow: Technically different from built-in distillation and still conditional on the exact DeepSeek license/model terms, Microsoft Product Terms, acceptable-use requirements, and the target model’s supported training options.
• “Open source” as sufficient permission: No. Open weights, hosted availability, inference access, and output-reuse rights are separate questions.

For a definitive approval, provide Microsoft Support with the exact model ID, version, deployment type, target/student model, whether outputs will leave Azure, and intended commercial use. Without those details—and an applicable license clause—a categorical “yes” is not supported.

References

1. Microsoft Foundry Models overview - Microsoft Foundry
2. Fine-tuning in Microsoft Foundry
3. Azure OpenAI stored completions & distillation (preview)
4. Model Distillation in Azure AI Foundry - Microsoft Q&A
5. Tutorial: Get started with a DeepSeek reasoning model in Foundry Models ...