Windows 11 ships Microsoft Execution Containers
At the October 7, 2026 event Microsoft gave AI agents their own sandboxed accounts and ran quantized models locally on 20 to 60 GB of RAM.
In short
On October 7, 2026 Microsoft made Microsoft Execution Containers generally available on Windows 11, giving each AI agent a sandbox and its own user account, and showed quantized models running locally in 20 to 60 GB of RAM on new laptops.
At a glance
- Microsoft Execution Containers (MXC) moved from preview to general availability on Windows 11.
- AI agents run as separate users with their own access limits; Microsoft Entra integration is announced, not shipped.
- MAI Code 1.1 Flash: 3-bit quantization, roughly 80 percent smaller, 53 GB, context up to 256,000 tokens.
- Nvidia Nemotron runs 2-bit quantized in 20 GB of RAM; DeepSeek V4 Flash uses 1.6-bit and 60 GB.
- Hybrid-intelligence laptops come from Asus, Dell, HP, Lenovo and MSI; the source lists no prices.
At its Windows and Surface October 2026 News event on October 7, 2026, Microsoft pushed Microsoft Execution Containers (MXC) out of preview and into general availability on Windows 11, so an AI agent runs inside a sandbox under its own user account with its own access limits. Satya Nadella and Windows chief Pavan Davuluri paired that with quantized models small enough to run locally in 20 to 60 GB of RAM.
The agent stops borrowing your permissions
This is the part that changes the threat model. Until now an agent acting on your desktop could reach whatever you could reach; with MXC it gets a separate account and a separate set of limits. Microsoft says the hook into Microsoft Entra, which would let an IT department manage those agent accounts centrally, is coming soon rather than being available today. Agent 365 is the layer meant to show what agents did across an organization.
Hybrid intelligence is an argument about the bill
Microsoft calls the split between on-device models and cloud models hybrid intelligence, and the pitch is cost rather than privacy. A developer demo burned well over one million tokens on the local machine while buying only a few thousand tokens externally. The Windows ML runtime spreads that work across GPU, NPU and CPU, and it now supports llama.cpp, which widens the set of open models developers can plug in.
What 20, 53 and 60 GB actually buy
Three models carried the local-execution demo. MAI Code 1.1 Flash is quantized to 3 bits, which cuts it by about 80 percent to 53 GB while keeping a context window of up to 256,000 tokens. Nvidia Nemotron is quantized to 2 bits and already fits in 20 GB of RAM, and DeepSeek V4 Flash runs at 1.6-bit quantization in 60 GB.
The practical reading for buyers: a fast NPU is not the constraint, memory is. Asus, Dell, HP, Lenovo and MSI are named as the hardware partners for hybrid intelligence.
Copilot Home, Copilot Code, Autopilot
The three Copilot features announced in late September got their place in the stack here. Copilot Home works from the local context of the PC, for example to tidy up files; Copilot Code uses hybrid intelligence to build native Windows apps; Autopilot takes on multi-step jobs, with assembling tax paperwork for an accountant as the example. Microsoft points to the coming months for Copilot+ PCs.
What this source does not establish
The hardware side stays thin: Microsoft showed laptops, desktops and servers that had been known for a while, and the first units can now actually be bought. The source gives no prices, processors, NPU figures or shipping dates, so none are claimed here. The Ars Technica report on the same event could not be retrieved while writing, which means everything above rests on the single heise online account and has not been cross-checked.
FAQ
What are Microsoft Execution Containers?
MXC is the sandboxed runtime for AI agents on Windows 11. It became generally available on October 7, 2026, and each agent gets its own user account with restricted access.
How much RAM do local AI models on Windows 11 need?
Based on the demos: 20 GB for Nvidia Nemotron at 2-bit, 53 GB for MAI Code 1.1 Flash at 3-bit, and 60 GB for DeepSeek V4 Flash at 1.6-bit.
When do Copilot Home, Copilot Code and Autopilot arrive?
Microsoft named the coming months for Copilot+ PCs and did not give a fixed release date for the three features.
