Microsoft announced on October 7 local time that it plans to make Windows the foundation for “hybrid intelligence.” The operating system will run AI agents on PCs when possible and use the cloud when more capability is needed. Windows and Devices Executive Vice President Pavan Davuluri explained the plan in an official blog post timed to Microsoft’s Windows and Surface event.

Microsoft said that as AI models grow larger, demand for AI use is beginning to outpace users’ cloud budgets. It said combining local and cloud computing can preserve leading-edge capabilities while reducing token consumption. For security, Microsoft began general availability of Microsoft Execution Containers (MXC), an execution environment that lets organizations set which files and networks agents can access and enforce policies while they run. Isolation options include process or session isolation, WSLc, virtual machines, and Windows 365 for Agents.

Microsoft said MXC management has three elements: containment, identity, and manageability. Integration with Microsoft Agent 365 and Intune is intended to govern agents across organizations. OpenAI Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA OpenShell, and Unsloth AI already support MXC; Claude Code, Box, Manus, Perplexity, and Raycast are expected to add support. Meta’s personal agent, Muse, is also expected to arrive soon as a Windows-native app integrated with MXC.

For local use, Microsoft is adding support for MAI Code 1.1 Flash, a coding model announced at Build in May. It has 137 billion total parameters and 6.8 billion active parameters. Microsoft says 3-bit quantization (compressing a model’s numerical values) reduces its size by about 80% while maintaining coding quality, and that it supports a 256,000-token (256K) context window (the amount of text a model can handle at once). The source says local operation appears to require a PC with RTX Spark or equivalent memory and computing resources. NVIDIA’s Nemotron, with more than 70 billion parameters, runs on about 20GB of memory using 2-bit quantization, while the 284-billion-parameter DeepSeek V4 Flash is also expected to run on an RTX Spark PC.

GitHub’s HydraFusion routes each task to the cloud model best suited to it. Microsoft is extending the technology to Windows and adding local models as routing options. Experimental previews are planned during October in the GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code. Windows ML, Microsoft’s model execution platform, is also adding support for the open-source inference engine llama.cpp.

On Copilot+ PCs, users will be able to let AI use on-device files and recent activity through “Local Context,” delegate file organization, device diagnostics, troubleshooting, and coding through “Local Actions,” and use on-device models for suitable tasks through “Local Models.” These capabilities will be applied to the new Copilot Home, Code, and Autopilot experiences. Code will be able to create Windows-native apps from a single prompt and run code securely using MXC. The persistent Autopilot will continue working while users do other tasks. Microsoft plans to roll out the features over the coming months.

Microsoft has opened reservations for the RTX Spark-equipped Surface Laptop Ultra. It has up to 128GB of unified memory and a 15-inch touchscreen, and Microsoft says it can run AI models with more than 120 billion parameters locally. Microsoft also announced the ASUS ProArt P16 and P14, Dell XPS 16 Creator Edition, HP OmniBook Ultra 16, Lenovo Yoga 9n 2-in-1, and MSI Prestige N16 Flip AI+; shipments begin October 16. The Surface RTX Spark Dev Box for developers is scheduled to ship in the United States in November. The Surface Laptop Ultra costs 513,480 yen for the 24GB model and 1,094,280 yen for the 128GB model.

Microsoft announced on October 7 local time that it plans to make Windows the foundation for “hybrid intelligence.” The operating system will run AI agents on PCs when possible and use the cloud when more capability is needed. Windows and Devices Executive Vice President Pavan Davuluri explained the plan in an official blog post timed to Microsoft’s Windows and Surface event.

Microsoft said that as AI models grow larger, demand for AI use is beginning to outpace users’ cloud budgets. It said combining local and cloud computing can preserve leading-edge capabilities while reducing token consumption. For security, Microsoft began general availability of Microsoft Execution Containers (MXC), an execution environment that lets organizations set which files and networks agents can access and enforce policies while they run. Isolation options include process or session isolation, WSLc, virtual machines, and Windows 365 for Agents.

Microsoft said MXC management has three elements: containment, identity, and manageability. Integration with Microsoft Agent 365 and Intune is intended to govern agents across organizations. OpenAI Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA OpenShell, and Unsloth AI already support MXC; Claude Code, Box, Manus, Perplexity, and Raycast are expected to add support. Meta’s personal agent, Muse, is also expected to arrive soon as a Windows-native app integrated with MXC.

For local use, Microsoft is adding support for MAI Code 1.1 Flash, a coding model announced at Build in May. It has 137 billion total parameters and 6.8 billion active parameters. Microsoft says 3-bit quantization (compressing a model’s numerical values) reduces its size by about 80% while maintaining coding quality, and that it supports a 256,000-token (256K) context window (the amount of text a model can handle at once). The source says local operation appears to require a PC with RTX Spark or equivalent memory and computing resources. NVIDIA’s Nemotron, with more than 70 billion parameters, runs on about 20GB of memory using 2-bit quantization, while the 284-billion-parameter DeepSeek V4 Flash is also expected to run on an RTX Spark PC.

GitHub’s HydraFusion routes each task to the cloud model best suited to it. Microsoft is extending the technology to Windows and adding local models as routing options. Experimental previews are planned during October in the GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code. Windows ML, Microsoft’s model execution platform, is also adding support for the open-source inference engine llama.cpp.

On Copilot+ PCs, users will be able to let AI use on-device files and recent activity through “Local Context,” delegate file organization, device diagnostics, troubleshooting, and coding through “Local Actions,” and use on-device models for suitable tasks through “Local Models.” These capabilities will be applied to the new Copilot Home, Code, and Autopilot experiences. Code will be able to create Windows-native apps from a single prompt and run code securely using MXC. The persistent Autopilot will continue working while users do other tasks. Microsoft plans to roll out the features over the coming months.

Microsoft has opened reservations for the RTX Spark-equipped Surface Laptop Ultra. It has up to 128GB of unified memory and a 15-inch touchscreen, and Microsoft says it can run AI models with more than 120 billion parameters locally. Microsoft also announced the ASUS ProArt P16 and P14, Dell XPS 16 Creator Edition, HP OmniBook Ultra 16, Lenovo Yoga 9n 2-in-1, and MSI Prestige N16 Flip AI+; shipments begin October 16. The Surface RTX Spark Dev Box for developers is scheduled to ship in the United States in November. The Surface Laptop Ultra costs 513,480 yen for the 24GB model and 1,094,280 yen for the 128GB model.