Sulat.com / Writing on AIThursday, November 19, 2026
Artificial Intelligence/18 November 2026

Microsoft will let companies customise its MAI models and run them in their own Azure

Microsoft says organisations can customise its MAI models and run them in their own Azure environments, keeping their data and changes. It's coming soon, with no price or model list yet.

Glossy 3D glass Microsoft logo beside an open steel vault in a server hall, with a glowing network sphere inside and a brass key on a console

At a glance

  • Microsoft says MAI Licensed Weights lets organisations deploy and customise its MAI models in their own Azure environments, with their data and adaptations remaining theirs and used to train Microsoft's models only with express agreement.
  • Microsoft says customers control how the models are deployed, who can access them and when they're upgraded. Axios reports the weights have to be deployed within Azure.
  • Microsoft cites AT&T, Woven by Toyota and Alaska Airlines. It says AT&T's customised model beat GPT-6.1 Sol on standard 5G tasks at one-fiftieth of the cost, and Alaska's classification accuracy rose from 81.2% to 91.1%.
  • MAI Licensed Weights is coming soon through Microsoft's Frontier Company engagements. Microsoft hasn't listed which models it covers, what the licence costs or when it starts.

Microsoft has introduced MAI Licensed Weights, a programme that lets organisations customise its MAI models and run them in their own Azure environments. Microsoft AI announced it on 17 November in the US, which was 18 November in New Zealand, and says it’s coming soon through Microsoft’s Frontier Company engagements. The weights are the numbers a model learns in training.

Microsoft AI framed it on X as something customers asked for. The Information reports that Microsoft unveiled a suite of new AI models at Ignite and will license their weights so customers can customise them.

What customers get

Microsoft says a customer’s data and adaptations remain theirs, and that Microsoft uses them to train its own models only with the customer’s express agreement. Customers also control how the models are deployed, who can access them and when they’re upgraded, which Microsoft says lets them move at their own pace.

Checklist: your data, your adaptations and your access list stay yours; Microsoft trains on them only with your agreement
Microsoft says the data, the tuned model and the access list all stay with the customer.

Microsoft’s Ignite post describes three steps: take a frontier MAI model, customise it on your own data if you want to, and run it in your own Azure environment. Axios reports that the weights have to be deployed within Azure. Microsoft’s Ignite blog says its trusted customers will get access to the weights, and Microsoft AI’s post says that happens through Frontier Company engagements.

Three steps: take a MAI model, customise it on your data, run it in your own Azure environment; coming soon via Frontier Company
The model is customised and run inside the customer’s own Azure environment.

What Microsoft says customers have done

Microsoft cites three customers, all tuning its Code 2 or Code 2 Flash coding models, which it released the same day. AT&T customised both on 5G network operations analytics. Microsoft says the customised model beat GPT-6.1 Sol on standard tasks at one-fiftieth of the cost, and reached frontier-level performance on harder ones at a significantly lower cost.

Woven by Toyota fine-tuned Code 2 Flash on historical MISRA coding-standard fixes. Microsoft says it produced ready-to-merge fixes in 44% of the cases where Woven’s existing frontier models needed a human, at less than a tenth of the cost. Alaska Airlines tuned Code 2 Flash to classify aircraft maintenance reports, and Microsoft says accuracy rose from 81.2% to 91.1%. Qualified employees keep responsibility for final decisions, Microsoft adds.

Stat cards: AT&T at 1/50th of GPT-6.1 Sol's cost, Woven by Toyota 44% merge-ready fixes, Alaska Airlines accuracy 81% to 91%
All three numbers are Microsoft’s, from tests each customer ran on its own task.

Microsoft’s footnote says the results come from customer-specific evaluations of models adapted for particular tasks, and that performance and cost advantages vary by workload, model configuration and evaluation method. The post doesn’t list the test sets or the dollar costs behind the fractions.

Microsoft hasn’t listed which MAI models the programme covers, what the licence costs or when it starts beyond “coming soon”. Companies that want access are told to contact their Microsoft account team.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.