Microsoft plans Windows AI to route work between PCs and the cloud

On October 7, 2026, Microsoft announced a Windows strategy designed to route AI work between a PC and cloud services according to the task. The plan brings together controls for AI agents, local models and cloud services; it does not mean every Copilot task runs locally. The announcement came after the Windows and Surface showcase covered in NeoTeo’s earlier preview.

Windows’ hybrid-AI strategy

In plain terms, Windows hybrid intelligence is Microsoft’s approach to placing AI work on a local PC or in the cloud, depending on what the task calls for. Microsoft says Windows should support both options rather than treat local and cloud computing as competing, all-or-nothing choices.

That distinction matters if you use a Windows PC for coding, writing or other AI-assisted work. A local model runs on the device; a cloud model runs through an online service. Microsoft says Windows will route work between them as appropriate, but its announcement does not describe a universal rule for deciding which model handles every request.

Local processing is therefore one part of Microsoft’s announced approach, not a promise that all AI work stays on the PC. Copilot’s planned features include local context, actions on the PC and access to local models, alongside the broader local-and-cloud strategy.

What MXC controls for AI agents

Microsoft says Microsoft Execution Containers (MXC) is generally available on Windows 11. MXC lets organizations set policies for which files and networks an AI agent can access, with those rules enforced at runtime.

Runtime enforcement means the access rules apply while an agent is operating. Microsoft also names process and session isolation, WSLc, virtual machines and Windows 365 for Agents among the Windows options for containing agent activity. The company says organizations managing governance at scale may need additional services.

For developers and IT teams, the practical point is that agent access can be governed through policies rather than left entirely to an agent’s intended behavior. MXC addresses what files and networks an agent may reach; it is separate from the question of whether a model runs locally or in the cloud.

Copilot and HydraFusion are the next layer

Microsoft plans to add local context, local actions and access to local models to Copilot on Copilot+ PCs. The company says Copilot will need permission to use local context from a PC. It expects the features to begin rolling out over the coming months, with availability varying by device, market and silicon platform. Microsoft names Copilot Home, Code and Autopilot among the experiences involved.

A separate plan concerns GitHub HydraFusion, a model-routing technology. Microsoft scheduled an experimental preview for later October 2026 that would extend routing in the GitHub Copilot app, Copilot CLI and Visual Studio Code to local Windows models as well as cloud models.

Windows ML and local models

Microsoft describes Windows ML as a runtime for deploying models across GPUs, neural processing units (NPUs) and CPUs, and announced support for llama.cpp, a framework used to run language models. The platform’s earlier release has a specific date: on September 23, 2025, Microsoft said Windows ML was generally available for production and included in Windows App SDK 1.8.1 for Windows 11 24H2 or newer. That is background on the runtime, distinct from the newer Copilot plans.

Microsoft also describes local model options, including MAI Code 1.1 Flash. The company gives the model 137 billion total parameters and 6.8 billion active parameters and describes a local 3-bit version with a 256K-token context window. It says the 3-bit version is nearly 80% smaller. These are Microsoft’s model specifications, not a measure of how well the model performs on a particular task.

Microsoft also names DeepSeek V4 Flash, a 284-billion-parameter model, as an option for local use on RTX Spark. An NVIDIA Nemotron model with more than 70 billion parameters is described as forthcoming; Microsoft says it uses 2-bit quantization and just over 20 GB of memory.

RTX Spark PCs and stated U.S. timing

Microsoft announced RTX Spark Windows PCs from ASUS, Dell, HP, Lenovo and MSI. It said the PCs would begin shipping in the United States on October 16, 2026. Microsoft also set October 16 as the start of Surface Laptop Ultra availability in the United States, and said Surface RTX Spark Dev Box would ship to U.S. customers in November 2026.

Microsoft describes Surface Laptop Ultra as having a 15-inch touchscreen and up to 128 GB of unified memory, and says it can run models exceeding 120 billion parameters locally. For a larger desktop system, Microsoft announced DGX Station for Windows, based on NVIDIA’s GB300 Grace Blackwell Ultra Desktop Superchip. Microsoft specifies up to 748 GB of coherent memory and 20 petaflops of FP4 AI compute, and describes workloads involving models up to 1 trillion parameters and teams running 32 or more agents simultaneously.

What Microsoft’s RTX Spark benchmark figures measure

Microsoft reported three “up to” results for selected preproduction RTX Spark Windows PCs against a preproduction 16-inch MacBook Pro with M5 Pro and 64 GB of memory. The testing was commissioned by Microsoft or NVIDIA, took place in September 2026 and used different workloads for each result. Microsoft says performance varies by device and configuration.

Each figure applies to its named workload and test setup, not to a single combined performance score. The results describe Microsoft- and NVIDIA-commissioned tests.

Windows Search and gaming plans

Microsoft also announced task actions in Windows Search, including changing settings, adjusting screen brightness or sound, arranging windows and sending a message. The company described an initial rollout for Windows Insiders in an experimental channel on October 7, 2026.

For gaming, Microsoft said Call of Duty would come to RTX Spark in 2027.