A glowing cloud security illustration shows protected data flowing from a vault to a network, with access controls for users.
Microsoft says it will let customers take its own MAI models, adapt them with proprietary data, and run the result inside their own Azure environments. The program is called MAI Licensed Weights. It was announced at Ignite on November 17, 2026. The Information's briefing reported it under the headline "Microsoft to Let Customers Customize Its AI Models' 'Weights'", but the text is paywalled. This article therefore relies on Microsoft's announcement and other coverage.

A glowing cloud security illustration shows protected data flowing from a vault to a network, with access controls for users. What Microsoft actually announced​

Microsoft describes MAI Licensed Weights as giving organizations the ability to deploy and customize Microsoft AI models in their own Azure environments. Microsoft's pitch is that the models customers run increasingly hold what makes the company distinctive. Customers therefore want control over who can access them, how they are adapted, what they cost to run, and when they are upgraded.

The data terms Microsoft states are:

  • Customer data and adaptations remain the customer's. Microsoft will use them to train its models only with express agreement.
  • Customers keep control over how the models are deployed, who can access them and when they are upgraded.

These are Microsoft's stated terms. The announcement does not include contract language.

Not self-service, and not open weights​

Availability is limited. Microsoft says MAI Licensed Weights is coming soon through its Frontier Company engagements. Interested customers are told to contact their Microsoft account team. GeekWire describes it as a way for companies to customize a Microsoft model with their own data and run it in their own Azure environment, with help from Microsoft's Frontier Company consulting arm. GeekWire also notes that AT&T used an earlier version of the program, Custom Models, to build a MAI model for 5G network analytics.

The "weights" in the name refers to the model's underlying parameters. TECHi explains that trusted customers will get access to the weights of Microsoft's models under the program. Microsoft's own blog says "trusted customers" will be able to get access to Microsoft AI model weights through Frontier Co. A Windows Forum summary of the keynote adds that access conditions weren't specified.

This is not a public download. Axios reports that Microsoft requires deployment within Azure. Specifically, Axios says Microsoft is offering enterprise customers licensed model weights that they can customize with proprietary data, while requiring deployment within Azure. Axios also quotes Microsoft AI CEO Mustafa Suleyman saying the approach would "feel like open source" while offering more security and control. That is an analogy, not an open-source licence.

The customer results, with Microsoft's caveats​

Microsoft backs the program with three customer examples:

  • AT&T: On 5G network operations analytics, Microsoft says MAI outperformed GPT-6.1 Sol at one-fiftieth of the cost on standard tasks. Microsoft's footnote says the customized models were MAI Code 2 Flash and Code 2. For more complex tasks, it claims frontier-level performance at significantly lower cost.
  • Woven by Toyota: Microsoft says MAI models produced ready-to-merge code fixes in 44% of cases where the company's existing frontier models required human intervention, at less than one-tenth of the cost. The footnote says the model was MAI Code 2 Flash, fine-tuned on historical MISRA coding-standard fixes.
  • Alaska Airlines: Tuning MAI Code 2 Flash for aircraft maintenance report classification raised accuracy from 81.2% to 91.1%. Microsoft says qualified employees remain responsible for final decisions.

Microsoft's own caveat applies to all three. The results reflect customer-specific evaluations of models adapted for particular tasks, and performance and cost advantages vary by workload, model configuration and evaluation method. No benchmark protocol or baseline pricing has been published. Treat the figures as vendor claims. Axios also cautions more broadly that Microsoft's performance comparisons rely on selected benchmarks, and some pricing advantages reflect promotional discounts. That remark is about the wider MAI launch, not these customer cases.

Why this matters for IT teams​

Until now, most enterprises have consumed frontier models as hosted services and tuned them lightly, if at all. This program points toward a different arrangement. The customer adapts a vendor's model on its own data and operates it in its own cloud tenancy. That could give more control over model behaviour and upgrade timing. It could also make rollouts less disruptive, since a model swapped out from under a production workflow is a familiar headache. Those benefits follow from Microsoft's description. They are not yet demonstrated.

The strategic angle is that Microsoft is selling its own models next to partners' models. Axios notes that Microsoft has long relied on OpenAI and other AI companies for its most advanced models, and describes the MAI push as a sign of how aggressively it will compete on price. The same Ignite stage also announced that OpenAI's GPT-6.2 is now available in Microsoft Foundry. Microsoft is hedging, and customers get more choice. Cheaper specialized models can also mean a trade-off. Axios notes that a model optimized for everyday coding won't match a general model on advanced math or scientific research.

Questions still open​

Microsoft's announcement does not answer several things an architect would need before planning around this:

  1. Which MAI models and customization methods qualify.
  2. Which Azure regions and compute configurations are supported.
  3. How licence and inference costs are calculated.
  4. Whether customers can export adapted weights, or hold them only inside Azure.
  5. How security isolation, evaluation, rollback and versioning work.
  6. What "remain yours" means in the governing contract.
  7. What qualifies a customer as "trusted", and what Frontier Company engagements involve and cost.

Bottom line​

MAI Licensed Weights is a roadmap item, not a product you can sign up for. The reasonable move for IT leaders with specialized, high-volume workloads is to raise it with their Microsoft account team. They should ask for pricing, contract terms and exportability in writing, and run their own evaluations. Microsoft's customer numbers are interesting, but they are Microsoft's own.

 

References

  1. Microsoft to Let Customers Customize Its AI Models’ ‘Weights’ - The Information The Information 2026-11-17T19:40:35+00:00
  2. Microsoft pushes its own AI models at Ignite; OpenAI is 10.8% of its revenue techi.com
  3. MAI Licensed Weights: Frontier performance you control | Microsoft AI microsoft.ai