Azure OpenAI Fine Tuning is Everywhere

If you’re building an AI agent and need to customize its behavior to specific domains, its interaction tone, or improve its tool selection, you should be fine tuning!

Our customers agree, but the challenge faced has been twofold: regional availability and the cost of experimentation.

Today, we’re bringing fine tuning of Azure OpenAI models to a dozen new AI Foundry regions with the public previews of Global Training and Developer Tier. With reduced pricing and global availability, AI Foundry makes fine tuning the latest OpenAI models on Azure more accessible and affordable than ever to bring your agents to life.

What is Global Training?

Global Training expands the reach of model customization with the affordable pricing of our other Global offerings:

• 🏋️‍♂️ Train the latest OpenAI models from over a dozen Azure OpenAI regions.
• 🤑 Save with lower per-token training rates than Standard training.

A screenshot of the Azure AI Foundry dialog for creating a new fine-tuned model.

Where is Global Training available?

Today, we’re launching Global Training in the following Azure OpenAI regions:

• Australia East
• Brazil South
• France Central
• Germany West Central
• Italy North
• Japan East (no vision support)
• Korea Central
• Norway East
• Poland Central
• Southeast Asia
• Spain Central
• South Africa North

We’ll be adding even more regions throughput the public preview and plan to make training truly global by general availability. We'll also add REST API support shortly after launch, so stay tuned.

What about Data Residency?

Keep in mind that while Azure maintains strict eyes-off data privacy policies, your training data and the resulting model weights may be copied to another Azure region.

If your compliance needs require data residency, no worries! You can continue to use the regional Standard training offering in North Central US, East US 2, or Sweden Central.

Wow! Is this a new Fine-Tuning experience I see?

We’ve also streamlined the fine-tuning wizard into a single screen. In addition to minimizing the number of hops to get a job started, we’ve also added the ability to automatically deploy the resulting model. Planning to run a post-training model evaluation? Save time and use the built-in automatic deployment of either Global or Developer!

Wait, what is Developer Tier?

Until today, validating a newly trained model meant leveraging AI Foundry’s production-grade inference services with the price tags to match. Why test in production if you don’t have to? 😉

Today, we’re also launching the public preview of the most affordable way for data scientists and developers to evaluate their customized OpenAI models: Developer Tier.

As part of the public preview, a Developer deployment lets you:

• Deploy fine-tuned GPT-4.1 and GPT-4.1-mini models from any training region or from Global Training with 24 hours of free fine-tuned model hosting.
• Accurately budget your testing costs by paying only per-token at the same rate as Global Standard deployments.
• Evaluate multiple models simultaneously to help choose which version to promote from development to production.

Just like our Global deployment type and Global Training, Developer does not provide data residency. If you need regional residency guarantees, the regional Standard deployment type is a better option.

How do I get started?

First, you’ll need a fine-tuned Azure OpenAI model from the GPT-4.1 family, specifically GPT4.1 and GPT4.1-mini. (Additional models coming soon!)

Screenshot of deploying a fine-tuned model as a Developer deployment.

You can then deploy your model like you would when using the current pay-as-you-go tiers via AI Foundry or the REST API. If you use the API, make sure to set the sku name to DeveloperTier. Remember that your TPM quota is shared across all concurrent Developer tier deployments within a subscription’s region, so set the capacity accordingly.

Once deployed, you have 24 hours to run inference against the model at which time it will auto-delete. You’re welcome to re-deploy and run additional testing, but to keep the capacity available for all customers we constrain the lifetime of your Developer deployment.

Any tips for using Developer Tier?

If you already have your own code-based test suite or a 3^rd party that uses the Chat Completions API, just point your OpenAI client to the Developer deployment. You can also use Evaluations within Foundry and select or create a new Developer deployment for use with a model candidate. A Developer deployment behaves exactly like a Global deployment, meaning your current tools just work. 😀