Skip to content
AI.info

The Pulse

Microsoft plans licensed MAI model weights through Frontier Company deals

Microsoft says organizations will be able to deploy and customize MAI models in their own Azure environments through a forthcoming licensed-weights offering. The company also reported customer-specific results from models adapted for network analysis, coding and aircraft maintenance classification.

Microsoft plans licensed MAI model weights through Frontier Company deals

Organizations will be able to deploy and customize Microsoft AI models in their own Azure environments under a licensed-weights offering the company says is coming soon. Microsoft introduced MAI Licensed Weights on Nov. 17 and said it would be offered through its Frontier Company engagements.

The offering is intended to let businesses adapt the models to their own knowledge and workflows. Microsoft directed organizations seeking more information to their account teams.

Control over deployment and data

Microsoft says customer data and model adaptations remain the customers’ own, and it will use them to train its models only with their express agreement. The company also says customers will control how the models are deployed, who can access them and when they are upgraded.

The announcement describes businesses incorporating proprietary knowledge and expertise into the models they run. It presents control over model upgrades as part of the offering for organizations that embed the models in their products and workflows.

Customer-specific tests

Microsoft reported results from models adapted for network operations, software code fixes and aircraft maintenance classification. The company cautioned that the results came from customer-specific evaluations and that performance and cost advantages vary with the workload, model configuration and evaluation method.

AT&T used customized MAI Code 2 Flash and MAI Code 2 for 5G network operations analytics. For standard tasks, Microsoft says the MAI models outperformed GPT-6.1 Sol at 1/50th of its cost; for more complex tasks, it says they achieved frontier-level performance at a significantly lower cost.

Woven by Toyota used MAI Code 2 Flash for software code fixes. According to Microsoft, it produced ready-to-merge fixes in 44% of cases where frontier models already used in production required human intervention, at less than one-tenth of the cost. The MAI model had been fine-tuned on historical fixes following the MISRA coding standard.

At Alaska Airlines, tuning MAI Code 2 Flash improved accuracy in classifying aircraft maintenance reports from 81.2% to 91.1%, Microsoft says. Qualified employees retain responsibility for final decisions.

Earlier model rollout

In June, Microsoft introduced a family of seven models developed in-house at Microsoft AI. The family covered image, voice, transcription, coding and reasoning tasks, including the coding model MAI Code 1 Flash.

That earlier announcement also described Microsoft Frontier Tuning, an approach using reinforcement learning environments through which MAI models can learn from an organization’s workflows. Microsoft said a tuned MAI model for Excel matched GPT 5.4 while being up to 10 times more efficient.

Microsoft said at the time that its models would be distributed through Foundry and made available to developers through OpenRouter, Fireworks and Baseten. It also said developers would be able to tune model weights themselves.

Sources

  1. Microsoft AI
  2. Microsoft AI

Related stories

Explore

More articles