Microsoft recasts Windows as 'hybrid intelligence' platform mixing local and cloud AI
In short
Microsoft said on October 7 (US time) that Windows will become a foundation for "hybrid intelligence", running AI agents on the PC where possible and using the cloud when more capability is needed. It made its Microsoft Execution Containers (MXC) agent sandbox generally available on Windows 11, said its MAI Code 1.1 Flash coding model can run locally after 3-bit quantization, and opened preorders for the RTX Spark-based Surface Laptop Ultra.
Read the full story 3 min read
Microsoft on October 7 (US time) set out a new direction that positions Windows as the foundation for "hybrid intelligence", in which AI agents run locally on the PC when possible and draw on the cloud when more capability is needed. The plan was presented at a Windows and Surface event in the United States, and Pavan Davuluri, executive vice president for Windows + Devices, explained it in an official blog post, ITmedia AI+ reported. According to Mynavi News, Microsoft CEO Satya Nadella described the effort at the start of the event as leading to a "next-generation Windows Agentic Platform". Microsoft said users' AI needs are starting to exceed cloud budgets as models grow, and that combining local and cloud AI keeps advanced capability while reducing token consumption. [ 1 , 2 , 3 , 4 ]
For security, Microsoft made Microsoft Execution Containers (MXC) generally available on Windows 11. Organisations define which files and networks an agent may access, and Windows enforces that policy at runtime. ITmedia AI+ reported that isolation options include process or session isolation, WSLc, virtual machines and Windows 365 for Agents. Agent actions can be distinguished from the user's own, and the system ties into Microsoft Agent 365 and Intune for organisation-wide control; Impress Watch noted that large-scale management may require additional services. OpenAI's Codex, GitHub Copilot, OpenClaw and Replit already support MXC, and Claude Code is due to follow. ITmedia AI+ also listed LM Studio, NVIDIA's OpenShell and Unsloth AI as supported, Box, Manus, Perplexity and Raycast as coming, and said Meta's personal AI agent Muse will soon arrive as a Windows native app integrated with MXC. [ 1 , 2 , 3 ]
On local models, Microsoft said its MAI Code 1.1 Flash coding model, with 137 billion total and 6.8 billion active parameters, has been quantized to 3 bits, reducing its size by about 80% while supporting a 256K-token context window. NVIDIA's Nemotron model of more than 70 billion parameters and the 284-billion-parameter DeepSeek V4 Flash will also run locally on RTX Spark PCs. Gigazine reported that running MAI-Code-1.1-Flash locally is recommended on PCs with 120GB or more of memory. Windows ML will add support for llama.cpp, and GitHub's HydraFusion routing will add local PC models as targets, with an experimental preview for the GitHub Copilot app, Copilot CLI and Visual Studio Code in October; Impress Watch said the preview starts in late October. [ 1 , 2 , 3 , 4 ]
Copilot on Copilot+ PCs will gain three local capabilities, local context, local actions and local models, allowing it, with the user's permission, to draw on files and recent activity on the PC and carry out tasks such as organising files. In Copilot's Code feature users can build Windows native apps from a prompt and run code through MXC, and the Autopilot agent can keep working while the user does something else. Mynavi News described a demo in which Autopilot gathered tax documents requested by an accountant into a folder and attached them to an email. Timing will vary by device, market and silicon platform. [ 1 , 2 , 3 , 4 ]
On hardware, Microsoft opened preorders for the Surface Laptop Ultra and the Surface RTX Spark Dev Box, both built on the RTX Spark chip it announced with NVIDIA in June. Impress Watch said the laptop has a 15-inch touch display and up to 128GB of unified memory and can run AI models of more than 120 billion parameters locally. [ 1 , 2 , 3 , 4 ]
Why it matters
The plan moves AI agents into the operating system itself, with Microsoft adding controls over which files and networks agents may touch and records that separate agent actions from the user's. The company said combining local and cloud models is meant to cut cloud token use and cost, which Japanese tech outlets presented as a shift in how Windows PCs will be used for AI work.
Key facts
- Microsoft announced on October 7 (US time) that Windows will be the foundation for "hybrid intelligence", combining local and cloud AI. [ 1 , 2 ]
- Microsoft Execution Containers (MXC), a runtime that enforces policies on which files and networks AI agents can access, is now generally available on Windows 11. [ 1 , 2 , 3 ]
- Agents already supporting MXC include OpenAI's Codex, GitHub Copilot and OpenClaw; Anthropic's Claude Code is due to follow. [ 1 , 2 , 3 ]
- The MAI Code 1.1 Flash coding model (137 billion total, 6.8 billion active parameters) was quantized to 3 bits, cutting its size by about 80%, and supports a 256K-token context locally. [ 1 , 2 , 3 ]
- GitHub's HydraFusion model-routing technology will be extended to include local models on the PC, with an experimental preview in October. [ 1 , 2 , 3 ]
- Windows ML will support the open-source inference engine llama.cpp. [ 1 , 2 , 3 ]
- Copilot on Copilot+ PCs will gain local context, local actions and local models, with user permission, rolling out over the coming months. [ 2 , 3 ]
- Preorders opened for the NVIDIA RTX Spark-based Surface Laptop Ultra, with up to 128GB of unified memory, and the Surface RTX Spark Dev Box. [ 1 , 2 , 3 ]
Confirmed by several sources
- Microsoft announced a strategy making Windows the foundation for hybrid intelligence that combines local and cloud AI. [ 1 , 2 , 3 , 4 ]
- Microsoft Execution Containers (MXC) are generally available, letting organisations set file and network policies for agents that are enforced at runtime. [ 1 , 2 , 3 ]
- OpenAI's Codex, GitHub Copilot and OpenClaw already support MXC. [ 1 , 2 , 3 ]
- Claude Code is due to support MXC in future. [ 1 , 3 ]
- MAI Code 1.1 Flash was quantized to 3 bits, cutting its size by about 80%, with a 256K-token context window locally. [ 1 , 2 , 3 ]
- Nemotron (over 70 billion parameters) and DeepSeek V4 Flash (284 billion parameters) will run locally on RTX Spark PCs. [ 1 , 2 ]
- GitHub's HydraFusion routing will be extended to local models on Windows. [ 1 , 2 , 3 ]
- Windows ML will support llama.cpp. [ 1 , 2 , 3 ]
- Copilot will use on-PC data only with the user's permission, and the new local Copilot features will roll out to Copilot+ PCs over the coming months. [ 1 , 2 , 3 , 4 ]
- Preorders opened for the RTX Spark-based Surface Laptop Ultra. [ 1 , 2 , 3 ]
Still unclear
- When MAI Code 1.1 Flash was first announced. ITmedia AI+ says it was announced at Microsoft's Build conference in May, while Gigazine presents it as newly announced and says its predecessor MAI-Code-1-Flash appeared in June 2026.
- Prices of the Surface Laptop Ultra and Surface RTX Spark Dev Box. Mynavi News says prices were announced but the documents do not give them.
- What hardware counts as "local" for running MAI Code 1.1 Flash. ITmedia AI+ says it presumably means RTX Spark PCs or machines with similar memory and compute; Gigazine reports a recommendation of 120GB or more of memory.
- Exact start date of the HydraFusion preview. ITmedia AI+ says during October; Impress Watch says from late October.
- Release or shipping date of the Surface Laptop Ultra. Impress Watch mentions October 16 but its text is cut off before saying what happens on that date.
What local media are saying
Timeline, local time
- ITmedia AI+ reports Microsoft's hybrid intelligence plan for Windows. [ 1 ]
- Mynavi News reports on the event, including Nadella's remarks and the Autopilot demo. [ 2 ]
- Impress Watch reports the plan and the Surface preorders. [ 3 ]
- Gigazine reports the MAI-Code-1.1-Flash release and the hybrid intelligence concept. [ 4 ]