AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

Microsoft will soon bring MAI model weights to companies’ own Azure environments

AI-generated contentWritten by AI

zdnet.com, microsoft.ai, the-decoder.com, drweb.de, thelatentdaily.com, pivotnews.ai · 7 sources · Sources published: 17–19 November 2026

At Ignite, Microsoft unveiled the MAI Licensed Weights programme, under which a company can fine-tune a Microsoft model on its own data and decide who uses the model and when updates are rolled out. The programme will be available soon, but Microsoft has not yet announced prices or which models are included.

Microsoft’s AI unit, Microsoft AI (MAI), announced the programme on 17 November at the Ignite conference. Model weights are the numerical values a model learns during training. A customer that holds the weights can train the model further rather than only sending queries to an API.

In its announcement, Microsoft said the customer decides who uses the model and when to deploy a new version. For a model accessed through an API, the provider changes the version. Data and adaptations remain with the customer, and Microsoft uses them to train its own models only with the customer’s express consent.

Even so, the models in the programme are not open-weight models. ZDNET reported that the weights cannot be downloaded from Hugging Face in the way Mistral’s Le Chonk can. Microsoft has not said whether the weights can be moved outside Azure, The Latent Daily reported.

Microsoft highlights three customer results. An MAI model customised by AT&T outperformed GPT-6.1 Sol on standard 5G network analysis tasks at one-fiftieth of its cost. At Alaska Airlines, accuracy in classifying maintenance reports rose from 81.2 per cent to 91.1 per cent after fine-tuning. A model fine-tuned by Woven by Toyota produced a fix that could be merged into the code as is in 44 per cent of the cases where the frontier models used in production needed human help. Microsoft itself says the results are customer-specific evaluations of models adapted to particular tasks, and that the benefits vary by workload. Germany’s Dr. Web points out that the test methods have not been published.

In his keynote, MAI chief executive Mustafa Suleyman made the case for MAI models by citing the high token prices of frontier models, and for licensing weights by saying that no one wants to pour their company’s know-how into someone else’s model. Suleyman also questioned whether the origins of open-weight models can be trusted.

Dr. Web points to European Commission guidelines. Under them, a company that modifies a general-purpose AI model is generally itself treated as the model’s provider if the modification uses more than a third of the compute of the original training. Ordinary fine-tuning will likely stay well below that threshold, the publication says. It advises companies to test the model on their own tasks and work out whether the licence and their own compute time cost less than token prices.

Microsoft says the programme will be available ‘soon’ through Frontier Company engagements, and interested companies should contact their Microsoft account team. Prices, licence terms, a release date and the models included in the programme have not been disclosed.

Report an error in this story