Microsoft vient de secouer le monde de l’IA : 7 nouveaux modèles d’IA

Microsoft just made its biggest move yet to prove it can compete at the AI frontier without leaning on open AI. For years, Microsoft's AI empire was built around other people's models. It poured billions into open AI, brought those models into Azure, pushed C-Pilot everywhere, and later backed Anthropic 2. But at Build 2026 in San Francisco, Microsoft changed the tone completely. It revealed seven in-house AI models, a new agent stack, an always personal agent called Scout, the Microsoft IQ intelligence layer, an AI security system for developers, and a major quantum chip upgrade with Myerana 2. This was Microsoft drawing a line, its own models, its own agents, its own intelligence layer, and eventually its own quantum path. The headline model is MAI thinking one Microsoft's first reasoning model trained from scratch. According to Microsoft, it was trained on clean commerciallylicicensed data without distillation from third party systems. That detail matters a lot because many companies are now trying to avoid legal and business risks around training data, model copying, and dependence on another company's frontier models. MAI Thinking One is described as a midsized reasoning model with 35 billion active parameters. Some reporting lists its context window at 256,000 tokens, while Microsoft's developer coverage also describes it with a 128,000 token context window. So, the exact public framing seems to vary depending on the release. And Microsoft is already making some pretty aggressive performance claims. Mustafa Sullean, the chief executive of Microsoft AI, said that after tuning its models for consulting firm McKenzie, Microsoft outperformed OpenAI's GPT 5.5 on quality while projecting around 10 times better cost efficiency based on public pricing data scaled across model sizes. That is a huge statement because Microsoft is basically saying that its own model stack can compete with the companies it invested in while also being much cheaper to run. Satya Nadella framed it very directly at the conference saying the time has come for every company to move from just consuming a frontier model to fully participating at the frontier. And that line basically sums up Microsoft's bigger strategy here. It does not want to only rent intelligence from open AI and anthropic forever. It wants to build intelligence, host it, sell it, optimize it, and control the economics of it inside Azure. The money side explains a lot here. Every time Microsoft uses thirdparty models, part of the cost goes to outside providers with its own models running on Azure, Microsoft controls the whole stack. That means lower costs, better margins, and cheaper AI tools for developers. Now, Microsoft also launched MAI code 1 Flash, a coding model designed to convert text descriptions into source code for apps and websites. This model is rolling out across GitHub, Copilot, and Visual Studio Code, which means Microsoft is not just releasing a model into some isolated lab environment. It is putting it directly into the developer tools millions of people already use. On the coding benchmark side, Microsoft says MI thinking 1 was preferred over Anthropics Claude Sonnet 4.6 in blind evaluations run by Serge, an independent human rating partner. The company also says it matches Claude Opus 4.6 on coding benchmarks including SWEBench Pro. According to developer coverage, the model is available now in private preview on Microsoft AI foundry and Microsoft also released a flash version designed for speed and efficiency. So the strategy is familiar, one stronger reasoning model for heavier work and smaller or faster variants for tasks where cost and speed matter more. The seven model family also expands beyond text encoding. Microsoft announced MAI image 2.5 and a flash variant supporting textto image and imageto image generation. That means users can describe what they want in plain language or pass in sketches and visuals to guide the generation. The model is already live in PowerPoint and is rolling out on one drive. Then there is MAI transcribe 1.5 which supports high accuracy transcription across 43 languages with streaming coming soon. Microsoft also introduced MAI voice 2 and its flash variant available in more than 15 additional languages and capable of producing new voice options. Add MAI code 1 for GitHub, Copilot, and VS Code and the picture becomes clear. Microsoft is building a broad AI model layer for office work, software development, voice, images, transcription, reasoning, and agents. Now, you've probably noticed how much attention Claude is getting right now. Anthropic keeps adding new models and features from Claude code and Claude artifacts to skills, connectors, design tools, and more. And honestly, it makes sense. Claude has become one of the most useful AI tools for turning an idea into something real. Whether that means building an app, creating a presentation, organizing research, planning your week, or speeding up work that would normally take a whole team. The problem is that a lot of people keep saying learn without actually showing you a clear way to use it properly. That's why today's sponsor is hosting the world's first Claude Aathon, a two-day live workshop happening this weekend from 10:00 a.m. to 700 p.m. Eastern time. It's a deep dive into Claude, practical use cases, and more than 10 other AI tools, and they're opening 1,000 free seats for a limited time. Inside the workshop, you'll learn how to use Claude for deep research, build artifacts and dashboards, create full presentations, set up connectors like Indeed for job search, build custom GPTs and agents, and use AI tools for visuals, videos, and automation. You'll also get bonus resources like claude codes, a prompt library, and a personalized AI toolkit builder. Link is in the description or scan the QR code to join before the free seats close. All right, now back to the video. And that brings us to Microsoft IQ. Microsoft IQ is now generally available and it is meant to be the unified intelligence layer that makes copilot and AI agents more aware of the actual organization they are working inside. The goal is to move agents away from generic chatbot behavior and connect them to business context, company data, and internal logic so they hallucinate less and act with more useful grounding. Inside Microsoft IQ, there are several parts. Work IQ captures how users work inside Microsoft 365. It understands people, emails, documents, meetings, organizational systems, external sources, and the relationships between them. Work IQ APIs are set to become available on June 16th, giving agents direct access to that kind of work data. Fabric IQ runs on Microsoft Fabric and works as a semantic foundation for structured business data. Microsoft describes it almost like an ontology, meaning it gives business data a more organized meaning layer. Foundry IQ then handles unstructured information, pulling from documents like wikis, policies, contracts, and even the live web. Microsoft also added web IQ, a new member of this family. Webq gives agents real world grounding through web search. It is model agnostic and native to the model context protocol which matters because MCP is becoming one of the main ways agents connect to external tools and data sources. Microsoft says web IQ returns relevant information blocks nearly two and a half times faster than the next best alternative. So Microsoft is building this stack from multiple angles. It has the models, it has the work context, it has the business data foundation, it has document retrieval and it has web grounding. Then it connects all of that to agents. One of the more interesting announcements here is Microsoft Scout, the company's first autopilot agent. Microsoft describes autopilots as a new category of always on agents. These agents stay active in the background, work autonomously, have their own identity, and act on your behalf under the permissions and policies you or your organization set. That own identity part is important. Scout does not operate as some anonymous shared service account. Every agent works under its own governed entra identity which means the actions it takes can be traced back to a known actor inside the company directory. Its credentials are scoped to the task, protected end to end, redacted from logs and diagnostics and managed like a firstparty Microsoft service. Scout is integrated across Microsoft 365 apps like Teams, Outlook, One Drive and Sharepoint. It connects to chats, email, calendar, contacts, documents, browser resources, local desktop resources, and MCP servers. You interact with it through Teams, and the desktop app extends its reach into the browser and local system. The practical idea is simple. Scout handles coordination work that normally piles up throughout the day. It can proactively schedule and coordinate meeting times across time zones, flag important meetings, generate prep materials, identify upcoming deliverables, automatically block calendar time, and spot risks such as stall decisions before they become bigger blockers. Microsoft says Scout learns over time through work IQ, building context around how you work, what you care about, and what needs to happen next. The company's own employees have already been using an early desktop experience and Microsoft is now expanding it to a select group of customers in private preview and Frontier organizations. Access requires Frontier enrollment intoune policy configuration and an opt-in at astation. Users with a GitHub copilot license can then download and install the experience. Scout is also built on OpenClaw, the open- source technology that showed up in November 2025 as a way to give agents a kind of always on operating rhythm. Microsoft is contributing policy conformance upstream to Open Claw so organizations running it can validate whether their environment meets security and compliance requirements and get an audit ready answer. And Microsoft is clearly aware that always agents raise serious enterprise security questions. Scout can only access resources and destinations that have been approved. Sensitive actions can require human signoff. Microsoft Purview policies, including sensitivity labels and data loss prevention rules, are enforced before anything is sent or written. So, the pitch is that Scout can act autonomously while still staying inside the organization's identity, access, security, and compliance structure. For developers, Microsoft also introduced code name Mdash, which is a pretty funny name because AI generated text is often full of M dashes. Under the name though, the product is serious. It is a multimodel agentic security system that deploys more than 100 agents to search for exploitable bugs in code. These agents reason about data flows, business logic, exploit chains, and contextaware fixes inside the developer portal. So, Microsoft is not just using agents for office work. It is also applying them to software security where a swarm of specialized agents can review code from different angles and try to find problems that a normal static scanner might miss. Then, Microsoft made another big announcement that moves outside normal AI software, Majorana 2. Majorana 2 is Microsoft's next generation topological quantum chip and the company says it was developed with help from Microsoft Discoveries Agentic AI. Microsoft claims the chip is 1,000 times more reliable than its previous generation of cubits. Its mean cubit lifetime is now 20 seconds with some instances lasting as long as 1 minute. That is a massive jump in quantum terms. Cubits are extremely fragile. Tiny changes in temperature, vibration, or environmental noise can knock them out of their quantum state. Many common approaches measure cubit lifetime in microsconds or milliseconds. Microsoft is saying Majorana 2 can hold its state for 20 seconds on average. The company compared the improvement to inventing a phone battery that would go from dying in a day to lasting nearly 3 years on a single charge. The analogy is dramatic yet it gives a decent sense of the scale Microsoft is claiming. Majorana 2 currently has just 12 cubits while a useful quantum computer would require millions. Still, Microsoft says the combination of reliability, 1 microcond operations, and very small cubit size around 1/100th of a millimeter puts it on a path toward a commercially valuable scalable quantum computer by 2029. Zulfi Alam, corporate vice president of Microsoft Quantum, said they expect to have a quantum machine in 2029 that can solve commercially viable reasonable problems. Microsoft's approach is based on topological cubits and Myurana based physics tied to a quasi particle first theorized in the 1930s by Italian physicist attoriana. This path has been controversial. Microsoft previously had to retract a 2018 Nature paper that claimed evidence for the particle, and the new chip and supporting research have not yet been peer-reviewed. Some physicists are asking for more information, so this is still an area where Microsoft's claims will face heavy scrutiny. The technical change behind Majorana 2 also includes a new material stack. Majora 1 used aluminum as a superconductor, while Majorana 2 uses lead. In this context, lead helps shield fragile cubits from cosmic disturbances that can make them unstable. Cetton Nyak, a Microsoft technical fellow, said the team needs to keep improving each year to reach a computer with major commercial and societal value and compared their progress to being 1,000 times better than last year. Microsoft Discovery is a big part of this story as well. The platform is now generally available and lets companies use teams of AI agents for frontier R and D. These agents can search, research, reason through problems, generate hypotheses, optimize experiments, and help validate ideas while human experts stay in control. Microsoft's own quantum team is already using it to manage workflows, automate measurements, improve fabrication, find hidden flaws, and propose better solutions. That is a big deal because the project has nearly two decades of data spread across different teams, formats, and systems. AI agents can connect that information, spot patterns, and track how software, chip design, materials, fabrication, physics, and measurements all affect each other. They also speed up experiments. Creating a topological quantum state means adjusting hundreds of parameters, and manual measurement can take weeks. Microsoft says Discovery helped build an AI agent that cut this cycle by orders of magnitude, adjusting voltages in parallel and mapping conditions continuously. One agent even found an unccalibrated temperature sensor hidden in fabrication data. And this all arrives at a very interesting moment in the broader AI market. Open AAI and Anthropic are both moving toward massive IPOs, while Microsoft is tied to both. It committed $13 billion to OpenAI, invested up to $5 billion in Anthropic, and sells both companies models through Azure. Now, Microsoft is their investor, partner, distributor, infrastructure provider, and direct competitor. That is why build 2026 feels bigger than a developer event. Microsoft wants more control over the AI stack from models and agents to developer tools, enterprise data, research platforms, and eventually quantum hardware. So now the real question is how fast Microsoft can turn all of this into products people actually use every day and how much pressure this puts on OpenAI, Anthropic, Google, and every other company trying to own the next layer of AI.