Microsoft is pushing more artificial-intelligence work from cloud data centres onto personal computers, pairing powerful local models with a new containment system intended to limit what autonomous agents can access and change.
The strategy, described as hybrid intelligence, allows Windows to route work between on-device models and cloud services. Microsoft says this can preserve access to advanced cloud capability while reducing token costs, improving responsiveness and keeping selected data processing local.
Local models become practical
Microsoft demonstrated a local version of its MAI Code 1.1 Flash coding model, which has 137 billion total parameters but activates 6.8 billion for a task. Using lower precision, the company says it has reduced the model’s size substantially while retaining a 256,000-token context window.
Windows is also adding llama.cpp support to Windows ML, allowing developers to run a broader range of open-source models across compatible CPUs, GPUs and neural processors. GitHub’s model-routing technology is expected to begin testing local execution in Copilot, its command-line interface and Visual Studio Code later in October.
This does not make the cloud obsolete. The aim is to use local compute for privacy-sensitive, repetitive or cost-conscious work while reserving cloud models for tasks that need greater capability.
The security problem follows the agent
Running an agent locally creates a direct risk: software that can write code, open files, use tools and act continuously may inherit too much of the signed-in user’s authority.
Microsoft Execution Containers, now generally available, create policy-defined boundaries outside the agent’s control. Developers or organisations can specify which files, network destinations, tools and user-interface resources a workload may use. The agent cannot grant itself wider access.
Different isolation levels are available, ranging from lightweight process containers to separate Windows sessions, Linux environments and experimental micro virtual machines. Microsoft says management policies will allow organisations to add restrictions through tools such as Intune.
Why this matters for enterprise IT
The most important shift is architectural. AI security is moving below the application layer and into operating-system identity, isolation and management. That gives IT teams a clearer way to distinguish human activity from agent activity and to apply least-privilege controls.
Adoption will still depend on hardware cost. High-memory PCs capable of running large models remain expensive, limiting early use to developers and specialist teams. Organisations must also test whether local models deliver sufficient accuracy for real work and whether container policies are usable at scale.
Even so, the direction is significant. If hybrid AI becomes standard, enterprises may gain more control over sensitive data and cloud spending—but endpoint management will become part of the AI platform. Security teams will need to govern not only which models employees use, but also what every local agent is permitted to do.