Microsoft Lets Businesses Fine-Tune MAI Models on Their Own Data in Azure
At Ignite, Microsoft launched Licensed Weights, a program that lets companies fine-tune MAI models on their own data and run them in Azure. It sits between OpenAI's and Anthropic's closed models and the open models people download from Hugging Face. The customer decides who gets access, and its data and fine-tuned versions stay with it. If you don't want to pour your company's knowledge into someone else's model, this option is worth raising with your Microsoft account manager.
AI-processed from ZDNet AI; edited by Hamidun News
At its Ignite conference, Microsoft showed several new models from its Microsoft AI (MAI) lab. But so many new models are coming out these days that the models alone hardly surprise anyone. What's more interesting is how Microsoft plans to deliver them to customers. The program is called Licensed Weights, and it addresses a question many companies are asking today: how to get a strong model without handing over your data and know-how to someone else.
What Licensed Weights is
The idea is simple. A company takes a frontier MAI model, fine-tunes it on its own data, and runs it inside Microsoft's Azure cloud. The model's weights, meaning the "settings" that hold everything it has learned, are not downloaded to the customer's servers. The customer instead gets its own customized version and control over it.
MAI head Mustafa Suleyman put it this way: you decide who has access to the model, and your data and modifications remain entirely yours. Microsoft says it will then work with the customer through its Frontier Company program to help deliver the custom model and integrate it into the customer's products and agents. Microsoft also said it will use customer information to train its own models only with the customer's consent.
You can start by contacting a Microsoft account manager. There is no public "connect" button yet, so in practice this is an enterprise deal that requires negotiation.
Where this sits on the open-versus-closed map
The market is usually described in simplified terms. Closed models like OpenAI's and Anthropic's are expensive and the strongest, but they can't be deeply tailored. Open models are cheaper and more flexible, but often weaker. Licensed Weights falls in the middle. It is more open than the APIs from OpenAI and Anthropic, but not open enough to post the model on Hugging Face, as Mistral did with Le Chonk.
Formally, this is not an open model family. Still, the approach adds nuance to the open-versus-closed debate and sets Microsoft's offering apart from competitors. On stage, Suleyman said no one wants to pour their intelligence into someone else's model. Palantir CEO Alex Karp has made a similar argument more than once.
The context: Western labs, including Thinking Machines and Mistral, are releasing increasingly competitive open models. But the bulk of the enterprise AI market still belongs to Anthropic and OpenAI. MAI models do not yet compete with them directly on capabilities. Microsoft does, however, have a huge base of enterprise customers, and that is its main lever.
Price and quality
Access to weights means little on its own if the model is weak. So Microsoft also released figures for its new coding model, MAI Code 2. According to Suleyman, averaged across four popular benchmarks of typical enterprise programming tasks, the kind found in GitHub Copilot, it delivers 98% of GPT 6.1 Sol's quality at 40% of the cost. Compared with Opus 5.5, the price is only 5%. These are Microsoft's own claims; the source contains no independent verification.
The company also released MAI Image 2.7, an image editing model. According to Artificial Analysis, it leads its category and beats Nano Banana Pro and several GPT models on price and quality.
What this means in practice
For product managers and founders, the takeaway is this. If you work with sensitive data, such as internal code, contracts, or customer records, Licensed Weights could be a compromise. You get a model trained on your material, and the cost is lower than the flagships from closed labs, if the stated numbers hold up.
There are honest limitations. The terms are not public yet: it is unknown what the program costs, what minimum volumes are required, or how deeply the model can be modified. The weights stay with Microsoft, so you are tied to Azure. Fully open models that you can run anywhere are also available.
A sensible step now: test MAI Code 2 against your current model on your own tasks, calculate the cost, and only then contact an account manager. If the quality is close to what's claimed and your data requirements are strict, Microsoft's option deserves serious consideration.
- Nov 21, 2026Microsoft outage: nearly 83,000 Outlook complaints on Downdetector, Teams and Copilot down too
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Microsoft Turns Copilot Into an "Operating System for Work": Word and Excel Become Agent Plugins
- Nov 21, 2026Google to split Gemini chat history from model training, so you can opt out without losing your history
Key Questions Answered
What Licensed Weights is?
The idea is simple. A company takes a frontier MAI model, fine-tunes it on its own data, and runs it inside Microsoft's Azure cloud. The model's weights, meaning the "settings" that hold everything it has learned, are not downloaded to the customer's servers. The customer instead gets its own customized version and control over it. MAI head Mustafa Suleyman put it this way: you decide who has access to the model, and your data and modifications remain entirely yours.
Where this sits on the open-versus-closed map?
The market is usually described in simplified terms. Closed models like OpenAI's and Anthropic's are expensive and the strongest, but they can't be deeply tailored. Open models are cheaper and more flexible, but often weaker. Licensed Weights falls in the middle. It is more open than the APIs from OpenAI and Anthropic, but not open enough to post the model on Hugging Face, as Mistral did with Le Chonk. Formally, this is not an open model family.
What are the key details on Price and quality?
Access to weights means little on its own if the model is weak. So Microsoft also released figures for its new coding model, MAI Code 2. According to Suleyman, averaged across four popular benchmarks of typical enterprise programming tasks, the kind found in GitHub Copilot, it delivers 98% of GPT 6.1 Sol's quality at 40% of the cost. Compared with Opus 5.5, the price is only 5%. These are Microsoft's own claims; the source contains no independent verification.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.