ML in Production · 4 min read

Microsoft's Licensed Weights lets firms tailor models in Azure

Licensed Weights lets enterprises adapt models with proprietary data in Azure. Microsoft also claims MAI Code 2 reaches 98% of GPT 6.1 Sol performance at 40% of cost.

By Tom Whitfield806 words

Key takeaways

  1. Microsoft announced Licensed Weights at the Tuesday Ignite 2026 keynote.
  2. Licensed Weights lets customers adapt frontier-model weights with their own data and run models securely in Azure.
  3. Microsoft claims MAI Code 2 reaches 98% of GPT 6.1 Sol performance at 40% of its cost.
  4. Microsoft claims MAI Code 2 costs 5% as much as Claude Opus 5.5.
  5. MAI Image 2.7, released Tuesday, led its category on Artificial Analysis, according to ZDNET.
Microsoft’s new Licensed Weights lets businesses customize AI models with their own data
Microsoft’s new Licensed Weights lets businesses customize AI models with their own data — Pexels / pixabay

By Tom Whitfield

Microsoft AI (MAI), the laboratory inside Microsoft, will let customers adapt frontier models, or high-end AI systems, with their own data and run the results securely in Azure, its cloud platform, through Licensed Weights.

At the Tuesday Ignite 2026 keynote, Microsoft CEO Satya Nadella announced several new MAI models. The release schedule has accelerated, although ZDNET described the event as one among a crowded wave of model launches.

What does Licensed Weights actually allow?

Licensed Weights offers selective model access, not full model ownership. Open-weight AI publishes numerical parameters for modification. Customers can customize weights, the numerical parameters governing model behavior, using proprietary data, while the resulting model remains hosted in Microsoft's Azure cloud under the announced terms.

It does not become freely downloadable from Hugging Face, unlike Mistral's Le Chonk. The offer therefore sits between closed systems from OpenAI and Anthropic and fully downloadable releases. Interested companies must engage their Microsoft account representatives; the announcement specifies no self-service route.

Microsoft AI CEO Mustafa Suleyman said: “You control who can access it. Your data and adaptations stay entirely yours.” He said Microsoft would work with customers through Frontier Company to deliver the custom model and embed it in customer products and agents.

Microsoft also said it would use customer information to train its models only with customer agreement. That condition puts data use on contractual consent, rather than an assumed product default. For teams protecting intellectual property, the decisive questions remain retention, access, and ownership of adaptations.

Why does the enterprise pitch matter?

The market narrative contrasts expensive, capable proprietary models with cheaper, adaptable open models that may perform less well, although ZDNET calls that division simplistic. Microsoft AI's models do not necessarily compete with OpenAI and Anthropic yet.

Microsoft does, however, retain legacy reach across countless businesses, according to ZDNET. Licensed Weights could turn that installed base into a controlled path toward customization without abandoning Azure. Western open-weight competition is rising, with Thinking Machines and Mistral releasing more competitive models. Yet ZDNET reported that OpenAI and Anthropic still command the overwhelming majority of the enterprise AI market.

Licensed Weights is not an open-weight family, but it broadens the closed-versus-open choice and differentiates Microsoft. The price-performance case still depends on buyers finding MAI's models impressive enough. The source lists no customer names, adoption count, or general-availability date.

What do the benchmark claims establish?

Suleyman said MAI Code 2 delivers “98% of the performance of GPT 6.1 Sol at just 40% of the cost.” He added that, against Claude Opus 5.5, MAI Code 2 costs 5% as much.

The company averaged four widely used benchmarks, standardized evaluations focused on everyday enterprise coding tasks observed in GitHub Copilot, its assistant product for software teams. This is a Microsoft claim, not an independently reported result. The source material provides no individual benchmark scores, benchmark names, pricing unit, response latency, or evaluation methodology.

That omission blocks independent calculation of both headline ratios. Data teams should request the underlying four tests, then reproduce them against internal repositories and coding workflows. Cost comparisons should include usage charges, retries, response latency, engineering time, and any Azure commitment.

What did MAI Image 2.7 add?

Microsoft released MAI Image 2.7 on Tuesday. The Artificial Analysis leaderboard places it first in its category, according to ZDNET. ZDNET also reported that it beats Nano Banana Pro and several GPT image models on price and quality.

The summary gives no price, quality metric, test set, or category definition, so buyers cannot reproduce the ranking from the announcement alone or verify its implied savings. Image teams should test representative editing workloads before treating the leaderboard position as a procurement result.

Microsoft also unveiled Copilot Workspaces at Ignite 2026; the source provides no specifications for that release. The broader announcements combine four claimed advantages: return on investment, customization, security controls, and capability.

What should data teams do next?

Start with a Microsoft account representative, because Licensed Weights has no announced self-service enrollment.

  • Treat the 98%, 40%, and 5% figures as vendor claims pending workload-level validation.
  • Request benchmark names, raw scores, pricing units, response latency, and workload assumptions for MAI Code 2.
  • Confirm in writing that customer data and adaptations remain the customer's and that model training requires explicit consent.
  • Run proprietary and open-weight candidates against the same internal evaluation set.
  • Price Azure hosting, migration, and vendor switching before accepting the 40% or 5% headline saving.
  • Compare MAI Image 2.7 with Nano Banana Pro and current GPT image tools using representative editing workloads.
  • Test access controls, deletion procedures, and audit logs; the benchmark ratios do not settle governance.

Licensed Weights offers a middle option, but it remains contract-bound and Azure-bound under the announced terms.

microsoft.ai (Original)

Price tags

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering regulation, policy and their impact on Analytics Digest is a professional daily at Analytics Digest.

16 articles

Read also

« Previous articleNext article »