[Digital Today reporter Chi-gyu Hwang] Microsoft is moving to develop Windows into a platform for AI agents. To do that, it put forward “hybrid intelligence” as a keyword, splitting AI work between PCs and the cloud depending on what fits best.

Pavan Davuluri (파반 다불루리), corporate vice president in charge of Microsoft Windows plus devices, said on the company’s official blog on Oct. 7 (local time) that it is making Windows a base for hybrid intelligence, and highlighted detailed strategies and points of differentiation.

Microsoft said it compressed its coding model “MAI Code 1.1 Flash,” unveiled at its May developer conference Build, cutting its size by nearly 80 percent so it can run on PCs. The model has 137 billion total parameters and 6.8 billion active parameters.

Surface Laptop Ultra equipped with Nvidia RTX Spark can run models with more than 120 billion parameters locally with up to 128GB unified memory. A Windows version of DGX Station due at year-end will be able to cover models with more than 1 trillion parameters. It can also run at least 32 agents at the same time and serve as a team-level “token factory.”

On hybrid intelligence, Microsoft is stressing cost.

The company said GitHub’s HydraFusion feature, which lets users choose the right AI model for each task, previously supported only cloud-based models. It now also covers models running on PCs, allowing tasks that can be completed on a PC to be handled there to save cloud costs.

Davuluri said, “As AI models grow, it has become difficult to cover it all with cloud budgets. Customers want to save tokens without giving up frontier-level performance.”

Security is also an area Microsoft sees as important.

MXC (Microsoft Execution Containers), which Microsoft officially launched on Windows 11, isolates agents by specifying which files and networks they can access and enforcing that during execution.

OpenAI Codex and GitHub Copilot have already moved to support MXC, and Meta’s personal AI agent Muse is also set to be introduced as a Windows app that runs on MXC. It also provides functions to identify what each agent did and to control agents via Agent 365 and Intune.

Davuluri said, “If inference stays local, organizations can keep sensitive data and intellectual property within their own environment.”

The company said a Copilot feature that, with user permission, reads files on a PC and recent work context, and handles file organization or troubleshooting on the user’s behalf, will be rolled out within a few months on Copilot Plus PCs.

HydraFusion local integration will be released as an experimental version in late October. The Windows version of DGX Station and an RTX Spark-based mini desktop will be introduced at year-end. A feature that runs tasks directly from the Windows search bar will also be available first only to users of the Windows Insider experimental channel for the time being.

For Microsoft, building an ecosystem around hybrid intelligence is also a task.

Several companies such as Anthropic Claude Code, Perplexity and Manus have signaled support for MXC, but it is still at the promise stage. Some also point out that because MXC supports other operating systems, Microsoft needs to better show why AI companies should choose Windows. Microsoft, for its part, is emphasizing that deep integration with Windows offers a broad range of options from process isolation to virtual machines.

Hardware could also be a variable. Running large models locally requires high-performance PCs with large unified memory. Microsoft said RTX Spark PCs are up to 2.1 times faster than Apple’s 16-inch MacBook Pro (M5 Pro) in first-token response speed, 4.3 times faster for AI image generation and 6.2 times faster for video generation, but all are based on “maximum” figures.