Run an AI conversation on your Mac after switching off Wi-Fi

A downloaded language model can keep answering on a Mac without an internet connection. LM Studio provides a desktop interface for loading that model and chatting with it locally. The useful test is not whether an old answer remains visible after Wi-Fi is switched off: it is whether the computer can generate a new one.[4]

This walkthrough follows LM Studio’s documentation, checked on 15 October 2026. It requires an Apple-silicon Mac and macOS 14 or later; Intel Macs are not supported. The documentation recommends 16 GB of memory, while noting that 8 GB can work with smaller models and modest context sizes. The exercise below is a proposed local-use check, not a hands-on result for a particular Mac or model.[1]

Set it up while still connected

1. Install LM Studio from its official website. Keep the connection available during setup and the first successful model load, so any required downloads can finish.
2. Open Discover and choose a model that the application indicates will fit your computer. For this first exercise, prefer a modest model rather than the largest available download.
3. Choose an appropriate downloadable model file. A quantised version uses a reduced-precision representation of the weights; the size and trade-off vary. The download page explains the available options.
4. Open Chat, use the model loader and select the downloaded model. Wait for it to load, then send a short message. A model file sitting on the disk is not yet a model ready to answer.

Loading allocates memory for the model’s weights and associated data. A model that fits as a file may still need additional working memory when it runs. For the first attempt, keep the document and conversation short rather than testing several limits at once.[1][2]

Now ask a question the model has not seen

Use the following invented notice. It contains all the facts needed for the exercise; no website, private record or current-world knowledge is required.

In a new chat, paste the notice and ask: “Turn this into a clear resident notice of no more than 70 words. Preserve every fact. Do not add a date, a reason for the interruption or reception opening hours.” Read the result against the original before changing anything else.

Next disconnect Wi-Fi and any other internet connection, including Ethernet. Keep the exercise to ordinary local chat without external tools. Ask a fresh follow-up: “Which details about reception are missing from the source?” A sound answer should identify missing information, such as opening hours, rather than invent it. This is a proposed check, not a test result claimed for any particular model.

LM Studio’s offline documentation says local chat can operate without an internet connection once the necessary files are present. Searching for models, downloading them and obtaining updates still require connectivity. A failure during setup is therefore a different problem from a loaded model being unable to generate a response.[4]

Check usefulness as well as independence

For this exercise, success has two parts. First, a new response arrives while the network is disconnected. Second, it handles the supplied facts correctly: Wednesday, the two-hour interval, the affected taps, reception and extension 204. The first checks local operation; the second checks whether the answer is any good.

If memory is tight, retry with a smaller model or a shorter context. If the answer adds unsupported details, revise the instruction and compare again. Neither step requires a new cloud service. The point is to establish one useful local workflow before deciding whether a larger model would genuinely improve it.[1][3]

Report an error or suggest a correction