Microsoft is introducing AI agents to Windows, allowing Copilot to access local files and user activity to provide context-aware assistance. These features will roll out to Copilot+ PCs in the coming months. Security for these agents is managed by Microsoft Execution Containers (MXC), which provide isolated environments with granular control over resource access.
The company is implementing a hybrid intelligence model, distributing computing tasks between local hardware and the cloud. New local models include the 137-billion parameter MAI Code 1.1 Flash, which is quantized to three bits, as well as Nvidia Nemotron (70 billion) and DeepSeek V4 (284 billion). Windows ML will support llama.cpp to optimize resource allocation across CPU, GPU, and NPU.
A new hardware category for developers and AI specialists features the Nvidia RTX Spark platform. Partners including Asus, Dell, HP, Lenovo, and MSI are launching devices based on this architecture. The Microsoft Surface Laptop Ultra, equipped with 128 GB of memory, is capable of running local models with up to 120 billion parameters.
High-end workstations using Nvidia DGX Station with GB300 Grace Blackwell Ultra chips will support models exceeding 1 trillion parameters, such as Llama 4 Maverick and DeepSeek V4 Pro. These systems can handle over 32 concurrent AI agents. Additionally, Windows search is being updated to perform direct system actions and provide AI-generated responses.
The Nvidia RTX Spark platform aims to enhance Arm-based gaming by supporting ray tracing and DirectStorage. Titles such as Fortnite and Alan Wake 2 are already available, with Call of Duty support expected in 2027. The platform also introduces the Autopilot personal AI agent for continuous task management.
