Microsoft's Licensed Weights Lets Firms Custom-Train Its AI Models
Microsoft's Licensed Weights lets enterprises custom-train MAI frontier models in Azure — a middle path between OpenAI's closed models and open weights, with cost claims that await independent proof.

Why it matters
- Microsoft announced Licensed Weights at its Ignite conference on Tuesday, letting customers customize MAI frontier model weights with their own data in Azure.
- MAI CEO Mustafa Suleyman said MAI Code 2 delivers 98% of GPT 6.1 Sol's performance at 40% of the cost, and 5% of Opus 5.5's cost.
- MAI Image 2.7, released Tuesday, leads the image editing category on Artificial Analysis, beating Nano Banana Pro and several GPT image models.
- Access to Licensed Weights requires going through Microsoft account representatives; it is not fully open-weight like Mistral's downloadable models.
- Microsoft said it will only use customer information to train its models with customer agreement.
Microsoft will let business customers customize the weights of its frontier AI models with their own data and run them securely in Azure, through a new initiative called Licensed Weights announced at the company's Ignite conference on Tuesday. The program, unveiled by Microsoft AI (MAI) — the company's in-house AI lab — is a selective take on open-weight AI: more customizable than proprietary models from OpenAI and Anthropic, but not freely downloadable from Hugging Face the way Mistral's models are. Companies that want in must go through their Microsoft account representatives, which immediately narrows who can participate.
The announcement came during a keynote in which Microsoft CEO Satya Nadella presented several new models from MAI. New model releases have become routine in the current market, and Microsoft released no shortage of them. The access model, not the models themselves, is what distinguishes this announcement.
"You control who can access it. Your data and adaptations stay entirely yours. And we'll work with you through Frontier Company to deliver the custom model and embed it in your own products and agents," Mustafa Suleyman, CEO of MAI, said during the keynote.
That quote carries a lot of promise, and each clause deserves scrutiny. Access control, data ownership, and hands-on delivery are exactly the assurances enterprises demand before handing proprietary data to a vendor. Microsoft has also stated it will only use customers' information to train its models if customers agree. That is a contractual commitment, not a technical guarantee, and the enforcement details live in agreements customers have not yet seen publicly.
What is Licensed Weights, exactly?
Weights are the numerical parameters inside a neural network that encode what a model has learned; customizing them means actually adapting the model itself, not just prompting it or bolting on retrieval. Under Licensed Weights, customers can adapt Microsoft's frontier models using their own data and deploy the results inside Azure's security perimeter.
The offering sits deliberately between two poles:
- Proprietary models from OpenAI and Anthropic: highest capability, no weight access.
- Open-weight models from labs like Mistral: fully downloadable, customizable, generally less capable on benchmarks.
Licensed Weights is neither. Microsoft grants weight-level customization but retains control over distribution — hence the account-rep gate. ZDNET, which first reported the announcement, noted that the initiative "isn't an open-weight family of models" but does diversify the open-versus-closed debate and differentiates Microsoft's enterprise offering from those of OpenAI and Anthropic.
The takeaway: Microsoft is selling the customization of open weights with the control surface of a proprietary vendor.
Why does this matter for the enterprise market?
The dominant, if simplified, narrative in AI holds that proprietary, expensive models deliver peak capability but resist customization, while open, cheaper models are customizable but less capable. Anthropic and OpenAI still hold the overwhelming majority of the enterprise AI market, according to ZDNET's reporting. Western labs including Thinking Machines and Mistral are releasing increasingly competitive open-weight models, but the shift is gradual.
Microsoft's own models do not necessarily compete with OpenAI and Anthropic on capability yet — a caveat the source itself makes. What Microsoft has instead is legacy sway: countless businesses already run on its software, cloud, and enterprise agreements. Offering a customizable AI option through existing channels could let Microsoft strike a balance among major tech companies that neither pure-play camps can match.
The appeal, per ZDNET, rests on three business priorities: privacy, capability, and cost. Licensed Weights addresses the first directly — adaptations and data stay with the customer. Capability depends on how the customized models actually perform in production. Cost is where Microsoft made its most concrete claim.
In short: Microsoft is betting distribution and control matter as much as raw benchmark scores.
What did the new models claim on performance?
Alongside Licensed Weights, Microsoft announced MAI Code 2, a coding model, and released MAI Image 2.7, an image editing model, on Tuesday.
Suleyman offered specific numbers for the coding model. "Averaged over four widely used benchmarks that target the everyday enterprise coding tasks we see in GitHub Copilot, MAI Code 2 gives you 98% of the performance of GPT 6.1 Sol at just 40% of the cost, and against Opus 5.5 it's an incredible 5% of that cost," he said.
Those figures are vendor-reported averages over unnamed "widely used benchmarks." A 98%-of-performance claim at 40% of cost is the kind of ratio that wins procurement conversations, but independent verification has not yet appeared, and averaged benchmark scores can obscure weak spots on specific tasks. The 5% cost figure versus Anthropic's Opus 5.5 is even more aggressive, and even more in need of outside testing.
MAI Image 2.7 now leads the image editing category on Artificial Analysis, an independent benchmarking tracker, according to the source, beating Nano Banana Pro and several GPT image models on both price and quality. That ranking, at least, comes from a third party rather than Microsoft's own marketing.
The takeaway: the price-performance claims are striking on paper and unverified off it.
What is not being said?
Several gaps remain. Microsoft did not publish pricing for Licensed Weights arrangements, the terms of the customization process, or which frontier models qualify. The "Frontier Company" delivery program Suleyman referenced has no detailed public description in the announcement. There is no stated timeline for general availability, and no named early customers.
The claim that customer data stays "entirely yours" also raises the question of what happens to adaptations at the end of a contract — whether custom weights can be exported, or whether they live only in Azure. Weight-level customization locked to one cloud is a meaningful lock-in vector, and enterprises evaluating the program will want that answered in writing.
Microsoft also framed the announcement partly through the open-weight trend. Western open-weight AI is trending upward, as ZDNET's takeaways note, and Microsoft clearly wants credit for openness while keeping distribution closed. Whether that positioning holds up depends on terms nobody outside the program has read.
The takeaway: the announcement answers who can customize the models, but not on what terms, at what price, or with what exit rights.
What comes next?
Microsoft also used Ignite to unveil Copilot Workspaces and to claim its new AI coding model rivals Claude Opus 5.5, signaling a broad portfolio push rather than a single product bet. With Licensed Weights, the company has positioned itself for organizations seeking a blend of ROI, customization, security controls, and capability — ZDNET's own assessment, and a fair one, provided businesses find the models impressive enough.
That proviso is the whole story. Microsoft's enterprise reach guarantees the program an audience; it does not guarantee that MAI's models close the capability gap with OpenAI and Anthropic, and the source itself concedes they are not there yet. If the customized models deliver, Licensed Weights becomes a genuine third path between closed frontier AI and open weights. If they do not, the account-rep gate and the unanswered pricing questions will make the program a hard sell.
Original: microsoft.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles
Related articles
- Mistral releases trillion-parameter ML4 weights for cyber defense
- Private AI Needs Infrastructure Built to Evolve, Not Just GPUs
- OpenAI Clears FedRAMP Moderate for ChatGPT Enterprise and Its API
- ServiceNow Makes OpenAI a Preferred Intelligence Layer for 80 Billion Workflows
- Microsoft’s AI Transparency Initiative Targets Enterprise Agents