On October 7, 2026, Microsoft announced plans to give Copilot on Copilot+ PCs permission-based access to local context, the ability to take actions across Windows and access to models running on the PC. Microsoft described the design as hybrid, with cloud intelligence retained when needed, and said rollout was expected in the coming months.
The announcement adds detail to the Windows-and-PC event covered in NeoTeo’s earlier report.
Microsoft plans local Copilot features for Copilot+ PCs
The three planned capabilities are local context, actions across Windows and access to local models. Microsoft said Copilot’s access to local context, including files and recent activity, would be permission-based. The announced scope is Copilot+ PCs, not every Windows 11 computer.
A Windows Search demonstration offers a concrete example of local file search and system actions:
Windows has three distinct routes to local AI
Windows AI APIs, Foundry Local and Windows ML are different tools for building or using local AI. Their roles, hardware support and model arrangements vary, as outlined in Microsoft’s Windows AI overview.
- Windows AI APIs give app developers ready-to-use functions for language, images, semantic search, speech, optical character recognition and video. Most target Copilot+ PCs; some support additional hardware.
- Foundry Local provides ready-to-use language and speech-to-text models. Microsoft handles model distribution and the runtime, and models can be shared across apps. Microsoft lists Windows 10 and later, while model availability and performance depend on the hardware.
- Windows ML is an ONNX Runtime-based framework for developers who want to run models they find or train. App developers handle model distribution, and compatibility and performance vary by device. Microsoft announced general availability for production use on September 23, 2025, and said Windows ML was included in Windows App SDK 1.8.1 for Windows 11 24H2 or later in its announcement.
The practical distinction is who chooses and supplies the model: Windows AI APIs provide ready-made capabilities, Foundry Local offers ready-to-use models, and Windows ML lets developers bring their own.
Hybrid AI keeps cloud processing in the picture
Local processing does not make Microsoft’s announced Copilot design entirely local. Microsoft said cloud intelligence would remain available when needed, while access to Copilot’s local context would require permission. The announcement describes a mix of local and cloud processing, rather than an offline mode for every task.
Microsoft Execution Containers set agent access policies
Microsoft said Microsoft Execution Containers (MXC) became generally available on Windows 11. Organizations can use MXC policies to specify which files and networks supported agents may access; those rules are enforced at runtime. That is a way to constrain an agent’s access, not a claim that all agents use the same policy.
What Microsoft said about MAI Code and Surface Laptop Ultra
Microsoft described MAI Code 1.1 Flash as a model with 137 billion total parameters and 6.8 billion active parameters. The company also cited 3-bit precision, a nearly 80% reduction in model size and a local context window of 256K tokens. These are Microsoft’s announced specifications.
The hardware figures are tied to specific products. Microsoft said Surface Laptop Ultra configurations would offer up to 128 GB of unified memory and local support for models exceeding 120 billion parameters. For RTX Spark systems, Microsoft also named a forthcoming Nemotron model with more than 70 billion parameters, 2-bit quantization and just over 20 GB of memory, alongside DeepSeek V4 Flash at 284 billion parameters.
HydraFusion’s planned Windows preview
Microsoft said it planned to extend HydraFusion to use models running locally on Windows. An experimental preview was planned for later in October 2026 in GitHub Copilot, GitHub Copilot CLI and Visual Studio Code.
