Direct answer: Microsoft says MAI-Code-1.1-Flash can now run locally, but this is not a lightweight model for an ordinary laptop. The local build has a 53GB model footprint, reached 75.5GB peak memory at the full 256K context in Microsoft’s test, and Microsoft recommends a PC with more than 120GB of RAM for best performance. GitHub Copilot’s automatic local/cloud routing and manual Windows ML selection are due experimentally by the end of October 2026—not necessarily available to every user today. [1] [2]
Important download-status note: Microsoft’s October 7 announcement says the model is “available to download and run locally,” but the announcement page did not provide a direct public model-artifact link or a standalone installation command when this guide was checked on October 8. Do not download similarly named files from an unofficial mirror. If the local option is absent, use MAI-Code-1.1-Flash in GitHub Copilot now and wait for the official Windows ML/Copilot rollout.
MAI-Code-1.1-Flash local: quick status table
| Question | Verified answer |
|---|---|
| Can it run on-device? | Yes. Microsoft announced an on-device, 3-bit-quantized build. |
| Is it already in GitHub Copilot? | The cloud model is in production in Copilot; the new local routing experience is rolling out separately. |
| Recommended memory | More than 120GB RAM for best performance. |
| Local model footprint | 53GB in Microsoft’s described quantized build. |
| Peak memory at 256K context | 75.5GB in Microsoft’s Surface Laptop Ultra test. |
| Context window | Up to 256K tokens. |
| Local inference charge | Microsoft says local model calls have zero inference charges. |
| Copilot local-model timing | Experimental access by the end of October 2026 across the Copilot app, CLI and VS Code. |
| Ordinary 16GB/32GB laptop? | No practical support claim has been published for those memory tiers. |
Microsoft describes MAI-Code-1.1-Flash as a mixture-of-experts coding model with 137 billion total parameters and 6.8 billion active parameters. Its local version uses quantization and speculative decoding to reduce memory use and improve responsiveness. [2]
Should you try it now?
Use it now in cloud Copilot if:
- You already use GitHub Copilot and want a fast coding model.
- Your PC has less than 120GB of usable memory.
- You do not need inference to remain on-device.
- You want the supported model-picker experience rather than a manual local runtime.
Prepare for local use if:
- You have a 128GB-class unified-memory workstation or similarly capable hardware.
- You want to reduce repeated cloud inference charges for eligible coding work.
- You need offline-capable inference for some tasks.
- You understand that local inference does not automatically make every Copilot tool call, MCP server or session fully offline.
That final distinction matters. Microsoft says model selection, inference and tool execution have different boundaries. A locally selected model can still be part of a session that uses networked tools or services, so “local model” should not be treated as a blanket privacy guarantee. [2]
How to check for MAI-Code-1.1-Flash in GitHub Copilot
- Update your Copilot surface. Update the GitHub Copilot app, Copilot CLI or Visual Studio Code and its Copilot extension.
- Open the model picker. Look first for MAI-Code-1.1-Flash. This may still be the cloud-hosted model.
- Check the provider label. For true on-device use, Microsoft says explicit selection will use MAI-Code-1.1-Flash through the Windows ML provider.
- Look for Auto routing. The upcoming HydraFusion-based experience can decide whether eligible work should run locally or use a cloud model.
- Test with a disposable repository. Start with a copied project, request a small code change, inspect the diff, run tests and confirm where inference occurred before using sensitive code.
- Enable sandboxing. In Copilot CLI, Microsoft documents the
/sandboxcommand for configuring sandbox settings. In the Copilot app, enable Sandbox new sessions for the project when available.
Do not assume “Auto” always means local. Microsoft says the router can use task context and cache state to choose between local and cloud inference. Manual selection is the better choice when location is a hard requirement. [2]
Hardware reality: 53GB is not the whole requirement
The compressed model footprint does not equal total memory required. The operating system, applications, inference runtime and key-value cache all need memory. Context growth also increases memory pressure as an agent reads files and receives tool results. In Microsoft’s first shipping Surface Laptop Ultra test, the model reached 75.5GB peak memory at the full 256K context; Microsoft separately recommends more than 120GB RAM for best performance. [1] [2]
| Your PC | Practical decision |
|---|---|
| 16GB or 32GB RAM | Use the cloud model. Do not expect the published local build to fit. |
| 64GB RAM | Below the published 75.5GB peak at full context; no official minimum-performance promise. |
| 96GB RAM | Potentially above the cited peak for one workload, but still below Microsoft’s best-performance recommendation. |
| 128GB unified memory | Closest to Microsoft’s demonstrated hardware class; usable memory is still shared with the OS and applications. |
This table is a decision aid based on Microsoft’s published measurements, not a new compatibility guarantee. Microsoft has not published a complete supported-device matrix or a hard minimum specification.
What is available now vs. coming later
Available or announced as available now
- MAI-Code-1.1-Flash in production in GitHub Copilot.
- Microsoft’s announced downloadable local build.
- Microsoft Execution Containers (MXC) generally available as a policy-driven containment layer for agent workloads.
- MXC support in agents including GitHub Copilot, OpenAI Codex, Replit, LM Studio and others.
MXC lets developers define file, network, process and interface boundaries outside the agent’s control. Microsoft documents process, session, WSL and micro-VM containment options, with availability varying by operating system and workload. [3]
Coming or rolling out
- Experimental local-model access across the GitHub Copilot app, CLI and VS Code by the end of October.
- HydraFusion routing between on-device and cloud models.
- Copilot hybrid-intelligence features on Copilot+ PCs over the coming months.
- Additional identity and management controls that distinguish agent activity from user activity.
Microsoft says Copilot’s local context, local actions and local-model capabilities are expected to begin rolling out on Copilot+ PCs in the coming months, with timing varying by device, market and silicon platform. [4]
Safety checklist before local agentic coding
- Use a copied repository or a clean Git branch for the first test.
- Keep production credentials outside the project directory.
- Grant write access only to the working folder.
- Block network access unless the task genuinely needs it.
- Review every diff before committing.
- Run project tests and security checks after generated changes.
- Verify whether MCP servers and shell tools are inside the sandbox boundary.
- Do not equate local inference with complete offline operation.
Microsoft says an agent’s shell commands normally inherit the permissions of the account that launches them. Sandboxing is therefore a separate control from local inference, not an optional synonym for it. [2]
Surface Laptop Ultra reference hardware
Microsoft demonstrated the local model on Surface Laptop Ultra hardware built around NVIDIA RTX Spark. The laptop can be configured with up to 128GB of unified memory, and Microsoft says it can run models exceeding 120 billion parameters locally. Pre-orders opened October 7 at a starting U.S. MSRP of $2,599, with availability beginning October 16; configurations and regional availability vary. [5]
You do not need to buy that specific laptop merely to use MAI-Code-1.1-Flash in cloud Copilot. Treat the Surface hardware as Microsoft’s reference class for the new local workflow—not as a requirement for the existing cloud model.
FAQ
Can I download MAI-Code-1.1-Flash now?
Microsoft says yes, but its October 7 announcement did not expose a direct standalone artifact link or install command when checked. Use only an official Microsoft, Windows ML or GitHub Copilot distribution path.
How much RAM does MAI-Code-1.1-Flash need locally?
Microsoft recommends more than 120GB for best performance. Its described quantized build is 53GB and peaked at 75.5GB at the full 256K context in one Surface Laptop Ultra test.
Will it run on a 64GB PC?
Microsoft has not published a 64GB support promise. The cited full-context peak exceeds 64GB, so the cloud model is the safer option unless Microsoft later publishes a smaller configuration or tested minimum.
Is local MAI-Code free?
Microsoft says local model calls have zero inference charges. That does not necessarily remove Copilot subscription requirements, hardware costs or charges for cloud models and connected services.
Does local inference keep all code offline?
Not automatically. The model can run locally while tools, remote MCP servers or router-selected cloud models use a network. Confirm the provider, sandbox and network policy for each session.
When will local selection appear in Copilot?
Microsoft says experimental access is planned by the end of October 2026 in the GitHub Copilot app, Copilot CLI and Visual Studio Code. Rollout timing can vary.
Sources
- Microsoft AI: MAI-Code-1.1-Flash
- Microsoft Command Line: local models and sandboxed tools in GitHub Copilot
- Windows Developer Blog: Microsoft Execution Containers
- Windows Experience Blog: Building Windows for hybrid intelligence
- Microsoft Devices Blog: Surface Laptop Ultra and RTX Spark Dev Box
Source check: October 8, 2026. Rollout details can change; verify the live Microsoft and GitHub Copilot interfaces before purchasing hardware.