Microsoft is turning AI into a portfolio of enterprise workflow models.

MAI is not one chatbot. It is a model family built around different kinds of work: reasoning through complex problems, helping developers write and change code, generating and editing images, transcribing audio, producing voice, and adapting to customer-specific processes.

That matters because enterprise AI is moving beyond generic demos. Real deployments need models that fit the workflow, run at the right cost, respect data boundaries, integrate into existing tools, and can be evaluated against business outcomes.

The practical buyer question changes. The old question was often, "Which model is best?" The better question is, "Which model is good enough, fast enough, governed enough, and integrated enough for this workflow?" MAI is Microsoft's answer to that more operational question.

Each model lane solves a different business problem.

Microsoft's June 2026 MAI release is not one general-purpose chatbot. It is a portfolio of capability lanes. Flash variants are the lower-cost, faster versions Microsoft describes for workloads where high volume and response speed matter.

The business claim is control: data, cost, integration, and tuning.

Microsoft says the new MAI models were developed in-house at Microsoft AI and built on a shared foundation with zero distillation. Distillation means training a model by having it imitate another model's answers. Microsoft says this family was not built by copying or imitating third-party model outputs.

Microsoft also emphasizes clean and appropriately licensed data. For MAI-Thinking-1 specifically, the technical report says pre-training used 30T tokens from publicly available and licensed human-generated data, avoided synthetic data generated by language models during pre-training, and made efforts to avoid and remove AI-generated content from collected sources.

Sparse MoE, short for sparse mixture of experts, means the model has many internal expert components but activates only some of them for each task. MAI-Thinking-1 is a 35B active / roughly 1T total parameter sparse MoE. In plain English, it has a large total capacity, but only part of that capacity turns on for any one request. That can improve efficiency because the system does not have to use the entire model for every task.

MAI-Thinking-1 is the reasoning proof point. MAI-Code-1-Flash is the developer workflow bet.

MAI-Thinking-1 is Microsoft's reasoning model for hard, multi-step tasks. Microsoft says it is a 35B active / 1T total parameter sparse MoE trained from scratch on 8K GB200 GPUs. The technical report lists 30T pre-training tokens, 3.55T mid-training tokens, and a 256K token context length. A token is a chunk of text a model processes. A 256K context window means the model can work over a very large amount of text in one request.

Microsoft reports 52.8% on SWE-Bench Pro, 97.0% on AIME 2025, 94.5% on AIME 2026, and 87.7% on LiveCodeBench v6. SWE-Bench Pro tests difficult real-world software engineering tasks. AIME is an advanced math competition benchmark. LiveCodeBench tests programming challenge performance. These are useful signals, but they are still vendor-reported benchmark numbers. Buyers should validate them on their own repositories, code review standards, and production constraints.

MAI-Code-1-Flash is aimed at everyday developer workflows. Microsoft says it was built end-to-end by Microsoft using clean and appropriately licensed data and is rolling out to GitHub Copilot individual users in Visual Studio Code through the model picker and Auto picker. Microsoft reports 51.2% versus 35.2% over Claude Haiku 4.5 on SWE-Bench Pro, and up to 60% fewer tokens on SWE-Bench Verified. Fewer tokens matter because they usually mean lower latency, lower compute cost, and faster interactive coding help.

The business value is not only text. It is the work around text.

Image, transcription, and voice models matter because many business workflows are not pure chat. Marketing teams need visual assets and edits. Meetings and call centers need searchable transcripts. Accessibility programs need high-quality speech. Voice agents need both accurate listening and natural speaking.

MAI-Image-2.5 supports image generation and image editing, with Microsoft also describing an ultra-efficient Flash variant. The enterprise question is whether the model can produce brand-safe, reusable creative work with acceptable review burden and predictable cost.

MAI-Transcribe-1.5 covers 43 languages on FLEURS, a multilingual speech benchmark. Microsoft reports best-in-class Word Error Rate, or WER, across those 43 languages. WER measures transcription mistakes, so lower is better. Microsoft also says the model can transcribe one hour of audio in under 15 seconds and includes keyword biasing, which lets teams supply important terms such as product names, acronyms, or medical vocabulary. Microsoft reports keyword biasing can reduce WER by up to 30% on FLEURS.

MAI-Voice-2 expands from English-only to 15 languages. Microsoft says it supports 5-60 second reference-audio voice prompting, includes built-in consent guardrails, is available in Azure Foundry, and is being integrated into Visual Studio Code and Dynamics 365 Contact Center. Microsoft also reports that MAI-Voice-2 was preferred over MAI-Voice-1 72% of the time.

Teach the model your company's private rules.

Generic models know a lot about the public internet. They do not automatically know how your company approves a deal, escalates a support case, reviews a pull request, manages a clinical workflow, writes a tax memo, or applies internal policy.

Frontier Tuning is Microsoft's approach for training or tuning a model around a customer's own workflows, inside the customer's environment. Microsoft describes reinforcement learning environments where the model learns from real workflows, tool usage, evaluation signals, company data, business processes, conventions, terminology, and access controls. Reinforcement learning means the model improves by trying actions and receiving feedback about whether those actions matched the desired behavior.

This can turn institutional knowledge into a private AI capability competitors cannot buy off the shelf. Microsoft says an MAI tuned model for Excel matches GPT 5.4 while being up to 10x more efficient. Microsoft also says a model tuned for McKinsey achieved the highest win rate of any tested model at roughly 10x lower cost. These are Microsoft examples that buyers should validate on their own workflows before treating them as procurement evidence.

Evaluate model-system fit, not model names.

• Use specialized models by workflow lane. Reasoning, coding, image, transcription, voice, and tuned enterprise workflows should each have separate evaluation criteria.
• Run workflow-specific evaluations. Test real repositories, meeting audio, domain vocabulary, design tasks, support scripts, latency targets, and error recovery paths instead of relying only on public leaderboards.
• Ask for written terms. Get clear answers on provenance, indemnification, customer-data use, third-party distillation, retention, logging, and human review obligations.
• Measure the full operating cost. Include latency, token cost, review burden, integration work, monitoring, governance work, and the cost of correcting model errors.
• Price the Flash variants separately. For high-volume use cases, a smaller, faster variant may create more business value than the most capable model.
• Treat Frontier Tuning as an operating model. It needs data owners, workflow owners, evaluation signals, security boundaries, model governance, and ongoing measurement. It is not a magic feature.

Sources used for this report.

This report uses Microsoft primary sources and treats Microsoft performance, preference, and cost figures as Microsoft-reported vendor claims unless independently validated.