Azure OpenAI Fine-Tuning
Microsoft AzureAzure OpenAI Fine-Tuning is Microsoft's managed fine-tuning service for OpenAI models inside Microsoft Foundry, with a preview tier that also tunes open-weight models. It supports supervised fine-tuning, direct preference optimization, and reinforcement fine-tuning, and wraps each job in Azure-native controls: VNet and private-endpoint isolation, RBAC, and content-safety filtering. Jobs are billed and deployed through Azure.
Openness was read from Microsoft's Product Terms, since the service publishes no license file; its reverse-engineering/weight-exfiltration restriction covers Foundry Models generally, not fine-tuning specifically. No usage figure for the feature alone is published, so adoption abstains. Verified 2026-08-09 via the Microsoft Foundry fine-tuning documentation, the Azure OpenAI data-privacy page, and Microsoft's Product Terms.
Openness
1 high confidence- source
- closed
- training-code
- closed
- runs-on
- Azure-AI-Foundry-only
- license
- Proprietary(managed Azure service)
A proprietary managed fine-tuning service inside Microsoft Foundry, with no source published. It runs only on Azure, and tunes both closed OpenAI models, which are generally available, and open-weight models in preview: Ministral, Qwen, Llama and gpt-oss.
- https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning recorded 2026-08-09
Supported-models table lists gpt-4o-mini as SFT-only; gpt-4o and the gpt-4.1/mini/nano trio as SFT+DPO; o4-mini and gpt-5 as RFT (gpt-5's RFT access gated by invitation); and public-preview SFT for Ministral-3B, Qwen-32B, Llama-3.3-70B-Instruct, and gpt-oss-20b. Fine-tuning runs only inside a Foundry project resource under the Foundry Owner/User RBAC roles; no repository or self-host path is offered anywhere on the page.
- https://www.microsoft.com/licensing/terms/product/ForOnlineServices/all recorded 2026-08-09
Defines "Covered Product" as "any Azure OpenAI model in Microsoft Foundry Models or any Copilot (excluding free Previews), in either case, that is available for a fee through Microsoft volume licensing or used with a paid subscription to an Online Service," and bars use of the Online Service "to reverse engineer any Online Service or exfiltrate the weights of any AI models or otherwise discover any underlying components, algorithms or systems included in the Online Service."
- https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy recorded 2026-08-09
"Your fine-tuned Models sold by Azure are available exclusively for your use." "Foundry is an Azure service; Microsoft hosts the Models sold by Azure in Microsoft's Azure environment and Models sold by Azure do NOT interact with any services operated by providers of Models sold by Azure, for example, OpenAI." Fine-tuned models ("base or fine-tuned") are deployed in the customer's Foundry resource and pass through the same real-time content-filtering pipeline as base models.
Adoption
not assessedNo standalone usage figure is disclosed for Azure OpenAI fine-tuning -- no jobs run, no count of customers tuning. Azure OpenAI is widely deployed across enterprises, but nothing published is specific to the fine-tuning offering, so no level is assigned rather than one borrowed from the umbrella product.
- https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning recorded 2026-08-09
The full article (supported models, methods, regions, RBAC prerequisites, file formats) carries no jobs-run, customer-count, or usage-volume figure anywhere; grepped the cleaned page text for customer/million/thousand/user/download/job-count/adoption terms — the only numeric hits are a joke example ("About 93 million miles") inside a sample training file and a recommendation to start with "50 well-crafted examples," neither a usage metric.
Capability
4 medium confidenceStrong managed method coverage over frontier closed base models, plus an open-model preview lane: supervised fine-tuning, DPO preference tuning and reinforcement fine-tuning with model-grader reward signals across the gpt-4o, gpt-4.1, o4-mini and gpt-5 line, with data validation, per-epoch checkpoints and auto-deploy, on Standard, Global and Developer training tiers. It is a black box -- no parallelism, precision or scale controls, and no training benchmark behind it -- and it sits level with the OpenAI fine-tuning API it wraps. It stops short of the top band because the user cannot control training internals or scale.
- https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning recorded 2026-08-09
Supported-models table: SFT only for gpt-4o-mini; SFT+DPO for gpt-4o and the gpt-4.1/mini/nano trio; RFT (GA) for o4-mini; RFT (GA, invite-gated) for gpt-5; public-preview SFT for Ministral-3B, Qwen-32B, Llama-3.3-70B-Instruct, and gpt-oss-20b. No parallelism, precision, or training-scale controls are exposed anywhere on the page.
Verified 2026-08-09