Microsoft's Local AI Push: What Hybrid Intelligence Means for Your PC

On October 7 in San Francisco, Microsoft laid out what it calls hybrid intelligence: Windows running AI agents and models on your own PC when that makes sense, and reaching the cloud for the hardest jobs. Satya Nadella shared the stage with Nvidia's Jensen Huang, and a new class of laptops built to run large AI models went up for pre-order the same day. Here's what changes for people who use AI at work, and what to check before you buy anything.

What is hybrid intelligence?

Microsoft describes Windows as "the home for hybrid intelligence: a platform where agents can run locally when it makes sense, reach the cloud when they need to," with security controls built in. For most people, the change will show up in Copilot. On Copilot+ PCs, the AI-focused Windows PCs Microsoft introduced in 2024, Copilot is getting three local abilities:

• Local context: with your permission, it can understand files and recent activity on your PC.
• Local actions: it can act for you across Windows, from organizing files to troubleshooting problems and running workflows.
• Local models: it can hand suitable tasks to AI models running on the PC and use the cloud when it needs to.

Jacob Andreou, Microsoft's executive vice president of Copilot, said Copilot will hand tasks to models on the PC "when cost or privacy matter more," keep the hardest tasks in the cloud, and act on your PC only with your permission.

In the stage demo, Copilot's Autopilot agent found an email from Andreou's accountant, gathered the tax documents scattered across his PC, sorted them into folders, zipped them and drafted a reply, then paused for his approval before sending. The sorting ran on the laptop, he said, so the tax details weren't sent to the cloud.

Why run AI on your own PC?

Cost is a big part of the pitch. Pavan Davuluri, who runs Windows and devices, wrote that "customers' needs are outpacing what their cloud budgets can support." Running models on the PC cuts the usage charges that come with cloud AI: a slide at the event said local models would run tasks "at no extra cost," and Nadella wrote on X that Microsoft is bringing "unmetered intelligence to every desk and every home."

Privacy is the other half. In the tax demo, the documents were sorted on the laptop instead of being uploaded.

There's a catch in the fine print. In a technical post written with GitHub, Microsoft notes that "local inference does not make the session offline." A model running on your laptop doesn't take everything else around it off the internet.

What can actually run on a PC?

The headline model is Microsoft's own MAI Code 1.1 Flash, a coding model with 137 billion parameters. Microsoft squeezed it to 53GB, about 80% smaller than the cloud version, and says it scores 70.8% on SWE-Bench Verified, a widely used coding benchmark, against 72.6% for the full model. Those are Microsoft's own tests. Microsoft is also bringing an upcoming Nvidia Nemotron model and DeepSeek V4 Flash, a 284-billion-parameter model, to its new RTX Spark machines.

The memory math matters. At its full 256K context window, the local coding model peaks at 75.5GB of memory on a Surface Laptop Ultra, according to Microsoft.

What hardware does it take?

There are two tiers. Copilot's new local features are expected on Copilot+ PCs, which Microsoft says now make up over 40% of laptops being built for business. The big local models are aimed at a new class of "builder" PCs running Nvidia's RTX Spark chip, from Asus, Dell, HP, Lenovo, MSI and Microsoft's Surface line.

The Surface Laptop Ultra, which Microsoft calls its most powerful Surface ever, starts at $2,599.99 with 24GB of memory and goes up to 128GB. It has a 15-inch touchscreen and ships October 16. That 24GB entry model can't fit the 53GB coding model in memory, so running it locally means a pricier configuration. A desktop version, the Surface RTX Spark Dev Box, costs $5,999 and ships in the U.S. in November.

Microsoft is pitching the machines against Apple. It says RTX Spark PCs return the first token of an AI response up to 2.1 times faster than a 16-inch MacBook Pro with M5 Pro, and generate AI images up to 4.3 times faster. Those figures come from tests Microsoft commissioned or Nvidia ran on preproduction machines. In the U.S. and Canada, Microsoft is also offering up to $1,000 back when you trade in an eligible MacBook Pro, through November 23.

How will Microsoft keep agents in check?

Davuluri was blunt about today's PC agents, calling them "fundamentally insecure because they have broad system access." Microsoft's answer is Microsoft Execution Containers (MXC), now generally available in Windows 11, which let organizations set which files and networks an agent can reach, enforced while it runs. On personal PCs, Microsoft says these safeguards are built into the agent experience itself.

OpenAI's Codex, GitHub Copilot, OpenClaw, Replit and LM Studio already support MXC, and Microsoft says Anthropic's Claude Code, Perplexity and others will follow. Meta's Muse agent is coming to Windows as a native app that uses it.

Microsoft has been here before. Recall, a feature that takes periodic screenshots of your activity so you can search it, drew objections from security researchers in 2024. Microsoft delayed it and made it opt-in before its wide release in 2025. This time, Microsoft says Copilot's access to your files and recent activity requires your permission.

When can you use it?

• Copilot's local features: expected to begin rolling out on Copilot+ PCs "in the coming months," though timing may vary by device, market and chip.
• GitHub Copilot: an experimental preview that sends simpler coding tasks to a local model and harder ones to the cloud arrives later in October.
• RTX Spark laptops: on pre-order now. The Surface Laptop Ultra ships October 16.
• Taskbar search actions, like "switch to dark mode" or texting a contact: rolling out to Windows Insiders now. Copilot answers in the search box will be opt-in and reach select markets later this year.

Fortune also noted that some of the features Microsoft showed aren't available yet.

Should you care?

If you use AI at work, yes, but you don't need to buy anything this month. The Copilot changes target Copilot+ PCs, and the RTX Spark machines, starting at $2,599.99 for the Surface, are built for developers and creators who want to run big models themselves. As this rolls out:

• When Copilot asks for access to your files and activity, think about what it actually needs before you say yes.
• Keep approval on for anything that sends, buys or deletes, the way the tax demo paused before sending.
• Don't read "local" as "offline." Check what each app still sends to the cloud.
• Treat the speed charts as marketing until independent reviews land.

My read: the laptops got the headlines, but the bigger change is where AI work happens. Microsoft is betting that a lot of everyday AI work, like sorting files or first-pass coding, doesn't need a data center, and I think that's right. Every task I hand to a cloud model counts against my plan, and plenty of those tasks are routine. Moving routine work onto the PC saves the cloud for the hard problems.

The security piece matters even more to me. I already let AI agents work on my files and in my browser, and what makes that safe enough to leave running overnight is rules: which tabs and apps an agent can reach, and what needs my approval. If Windows enforces those limits at the system level, and Claude Code adds MXC support as Microsoft says Anthropic plans to, agents on Windows get safer whichever one you use.

What I'd wait for: real-world reviews and the actual rollout. Copilot's local features are still months away, and the record-setting numbers come from Microsoft's and Nvidia's own tests. How are you splitting work between local and cloud AI? Tell me in the comments.

Quick answers

Do I need a new PC? Not if you already have a Copilot+ PC: that's where Microsoft expects Copilot's new features to start rolling out in the coming months, though timing may vary by device, market and chip. Running large models like MAI Code 1.1 Flash takes an RTX Spark-class machine with plenty of memory.

Will Copilot read my files without asking? Microsoft says no: Copilot needs your permission to use your files and recent activity, and it acts on your PC only with your permission.

Does "local" mean offline? No. Microsoft's own post with GitHub says local inference does not make the session offline.

How much is the Surface Laptop Ultra? It starts at $2,599.99 with 24GB of memory, goes up to 128GB, and ships October 16.

Sources

Primary sources

• Microsoft Windows Experience Blog: Building Windows for hybrid intelligence, Oct 7, 2026
• Microsoft Command Line: Bringing local models and sandboxed tools to Windows and GitHub Copilot, Oct 7, 2026
• Microsoft Devices Blog: Pre-order our most powerful Surface devices ever, Oct 7, 2026
• Microsoft: Surface Laptop Ultra

Reporting

• GeekWire: Microsoft plays the Windows card in the AI game, with help from Nvidia, Oct 7, 2026
• Fortune: Microsoft shows off new Windows software, revamped for agentic AI, Oct 7, 2026

Photos

• Cover, hybrid intelligence model and MXC partners: Microsoft, Windows Experience Blog
• Satya Nadella and Jensen Huang: NVIDIA, NVIDIA Blog
• Surface Laptop Ultra and Dev Box: Microsoft, Microsoft Devices Blog