Microsoft is introducing powerful new updates to Azure AI Foundry, giving organizations three advanced options to fine-tune AI models tailored to their unique business needs.

These latest enhancements include.

• Reinforcement Fine-Tuning (RFT) for the upcoming o4-mini model
• Supervised Fine-Tuning (SFT) for the newly available GPT-4.1-nano model
• Fine-tuning support for the powerful open-source Llama 4 Scout model

These improvements show Microsoft’s ongoing commitment to helping companies build smarter, more tailored AI systems that can make a real difference in the real world.

What’s New?
 

1. Reinforcement Fine-Tuning (RFT) with o4-mini – Coming Soon

RFT is a new way to train AI models by rewarding good behavior and penalizing wrong outputs. It helps AI models make better decisions, especially in complex or high-stakes environments. This is especially useful when your business logic is too detailed or unique for basic training methods.

The o4-mini model will soon support RFT, offering fast performance with better reasoning. It’s a great match for industries needing adaptive AI, such as finance, law, and healthcare.

Real-World Example: DraftWise

Legal tech startup DraftWise used RFT on Azure AI Foundry to improve how its models review and generate legal contracts. By training with their own legal data, DraftWise improved the model's response accuracy by 30%, helping lawyers work faster and smarter.

RFT helps our models understand legal language better and respond more intelligently to complex instructions,” said James Ding, CEO of DraftWise.

When should you use RFT?

RFT is best when,

• You need custom rules in your AI.
• You work in a specialized field with unique processes.
• You deal with complex decision-making.

Example Scenario: Contoso Wellness

This fictional wealth management company used RFT to train a model that follows specific rules, like identifying ideal clients based on net worth. This made onboarding faster and more efficient.

2. Supervised Fine-Tuning for GPT-4.1-nano – Available Now

Supervised Fine-Tuning (SFT) lets you teach a model using labeled examples so it learns exactly how your organization talks, works and delivers results.

Why use GPT-4.1-nano?

• It’s small, fast, and cost-effective.
• Great for customer service bots, internal tools, or document parsing.
• Keeps up with heavy workloads while maintaining high accuracy.

You can fine-tune it to,

• Match your brand’s tone and terminology
• Produce structured outputs
• Handle thousands of requests quickly and consistently

Bonus: It’s also perfect for distillation training it using data generated by bigger models like GPT-4.1 or o4 to keep costs down while maintaining quality.

3. Fine-Tuning Llama 4 Scout – Available Now

Meta’s Llama 4 Scout model is now supported for fine-tuning in Azure AI Foundry. It’s a top-tier open-source model with 17 billion parameters and a huge 10 million-token context window. Best of all, it fits on a single H100 GPU for fast performance.

You can fine-tune it using your own GPUs, and it’s available in,

• Azure AI Foundry
• Azure Machine Learning

With access to extra tuning options, developers have more control over how the model behaves making it even more customizable than the default serverless setup.

Build the Future of AI with Azure AI Foundry

These updates give organizations more power and flexibility to create AI that is,

• Trusted
• Efficient
• Customized to your exact needs

Whether you need,

• Adaptive reasoning (with o4-mini and RFT)
• Fast, low-cost intelligence (with GPT-4.1-nano and SFT)
• Advanced open-source customization (with Llama 4 Scout)

Azure AI Foundry has the tools to help.

Start building today

• Try the new fine-tuning features in Azure AI Foundry
• Use Azure Machine Learning for even more control
• Check out Microsoft’s documentation and model catalog

Microsoft says this is just the beginning. More models, fine-tuning tools, and features are coming soon—helping you create smarter, safer, and more personalized AI.