Running a Model Locally Is Not the Same as Keeping It Local
A local model and an offline application are different things. Three checks for LM Studio, what a Windows Firewall test showed, and the limits of that result.
It is easy to assume that a local LLM means nothing you type can leave your computer.
Two questions need separate answers: where does the AI model run, and what network connections can the application make?
Local applications can still have network features
LM Studio can run models without an internet connection once the model and its required runtime are available on your computer.
It also has features that use a network: searching for and downloading models, serving requests from other devices, and connecting to external services.
Using a local LLM does not by itself mean all network traffic has stopped. The presence of network features alone does not establish that your prompts are being sent out, either.
Start with these three checks
You do not need to understand every setting to begin:
- Can you use the model without a network connection after downloading it?
- Is the application configured to accept connections from other computers?
- Are you using connections to external services or additional tools?
LM Studio can serve only the same computer or allow access from other devices on your local network.
If you want to use it only on your own computer, check whether Serve on Local Network is enabled in the server settings. This allows access over the same LAN; it is distinct from publishing a server on the internet.
Check what happens, as well as the settings
In a test on 20 August 2026, we restricted external traffic associated with LM Studio using Windows Firewall. We then tried model searches, update checks, and related operations. Those requests failed on the one machine and configuration tested.
The useful step was trying the connections after applying the settings.
That result covers the LM Studio traffic we targeted. It does not mean Windows or other applications stopped communicating. Restrictions also need checking again after updates to the application or runtime.
Check where processing happens and which connections remain possible. That gives a clearer picture than relying on the word “local.”
More detail: the test and the published scripts
The test used Windows Firewall to restrict LM Studio traffic. It did not use a sandbox or virtual machine.
The configuration and test scripts are published in lmstudio-windows-hardening.
The result in this article is based on the operator’s report from 20 August 2026. It does not cover every Windows or LM Studio version, nor was it a comprehensive audit of possible routes for transmitting prompts.
In the published version checked on 13 September 2026, creating and auditing the project’s own firewall rules is disabled by default. The deployment setting ProjectFirewall = 'ON' enables that mode. The default, OFF, delegates enforcement to organizational controls; it does not mean the script has verified a network block.
The published version remains a preview, and its README says the revised shared-folder workflow requires a complete retest. The earlier result from one configuration does not validate the entire current distribution.
Sources
- LM Studio Documentation: Offline Operation
- LM Studio Documentation: Serve on Local Network
- LM Studio Documentation: Server Settings
- Microsoft Learn: Windows Firewall
Sources checked on 13 September 2026.