Local Microsoft LLMs + GitHub Copilot + Windows sandboxing

🚀 I saw it… I wanted to try it… but my PC isn’t powerful enough 😅

Recently Microsoft published something huge for developers: local large language models running directly on Windows, with sandboxed tool execution, integrated into GitHub Copilot.

This is not “AI in the cloud”. This is AI on your machine — offline, private, fast, and capable of running real automations.


🧠 What Microsoft is shipping

1) Local inference inside GitHub Copilot

Copilot can now automatically choose:

  • local model (on your PC), or

  • cloud model (Azure-scale).

Or you can force “local only”.

2) MAI Code 1.1 Flash — Microsoft’s local coding model

  • 137B total parameters

  • 6.8B active (Mixture‑of‑Experts)

  • 53 GB quantized size

  • Up to 923 tokens/sec

  • Up to 256k context

  • Optimized for coding + tool use

📌 Official model release: https://github.com/microsoft/MAI-Model-Explorer (github.com in Bing) https://huggingface.co/microsoft/MAI-Code-1.1-Flash (huggingface.co in Bing)

📌 Microsoft blog: https://aka.ms/MAI-Code-Flash (aka.ms in Bing)

3) MXC — Microsoft Execution Containers (sandboxing)

This is the key security layer.

MXC creates a sandbox around any command the AI runs:

  • restricted file access

  • restricted network access

  • no credential access

  • no system API access

  • no ability to delete or exfiltrate data

📌 MXC repo: https://github.com/microsoft/microsoft-execution-containers (github.com in Bing)

📌 Windows sandboxing explanation: https://learn.microsoft.com/windows/security/threat-protection/windows-sandbox/windows-sandbox-overview (learn.microsoft.com in Bing)

🛠️ What you can actually do with it

With local models + sandboxing, you can:

✔ Work fully offline

Write code, analyze repos, generate scripts — no cloud needed.

✔ Run automations

Daily dashboards, test runners, file processors.

✔ Execute commands safely

Copilot can run PowerShell/Bash inside a sandbox.

✔ Build local agents

Your own “mini‑Copilot” that:

  • reads your repo

  • generates HTML reports

  • runs tests

  • updates JSON history

  • refreshes dashboards daily

📌 Copilot CLI: https://github.com/github/copilot-cli (github.com in Bing)

📌 Copilot App: https://github.com/github/copilot-app (github.com in Bing)

🧰 Software you need

To try Microsoft’s local LLM setup:

Optional:

🔥 Hardware requirements (the painful part)

MAI Code 1.1 Flash is huge.

To run it locally you need:

  • 64–128 GB RAM

  • fast SSD

  • modern GPU or NPU

❓ AMD Ryzen AI — supported?

Yes, through DirectML.

Microsoft officially supports:

  • AMD Ryzen AI NPUs

  • AMD Radeon GPUs

📌 DirectML hardware support: https://learn.microsoft.com/windows/ai/directml/gpu-support (learn.microsoft.com in Bing)

But:

  • performance is lower than NVIDIA RTX Spark

  • still requires massive RAM

  • speculative decoding may be slower

If you have a normal laptop with 16–32 GB RAM → MAI Code 1.1 Flash will not fit.

🔐 Security: can the AI delete files or steal data?

Short answer: No — if MXC sandboxing is enabled.

MXC prevents:

  • deleting system files

  • reading sensitive folders

  • sending data to the internet

  • accessing credentials

  • running dangerous commands

Even if the model generates a harmful command, MXC blocks it automatically.

📌 MXC security details: https://github.com/microsoft/microsoft-execution-containers#security (github.com in Bing)

📌 Copilot sandbox documentation: https://commandline.microsoft.com/local-models-sandboxed-tools-github-windows/

🎯 Final thoughts

Microsoft is building a future where:

  • AI runs locally,

  • tools run in secure sandboxes,

  • developers get cloud‑level intelligence on their own hardware.

I wanted to try MAI Code 1.1 Flash myself… but my PC doesn’t have 128 GB RAM 😭 So for now — I’m watching from the sidelines and experimenting with smaller models.

But the direction is clear: local AI is becoming real.