A model on your own laptop

This mini-book is for Level 1 User. You are running one model for your own work, on your own laptop, through a desktop application. If it is slow, wrong, or unavailable, only your work should be affected. You are not creating a service for colleagues or connecting the model to other tools.

2. The cloud tool is not an option

You need to turn rough notes into a short private draft. In a lab, the notes resemble observations from an unpublished experiment. In a company, they resemble notes about a confidential staffing matter. The cloud assistant you normally use is not approved for either kind of material.

You could abandon the useful drafting task, or you could run a small model directly on your laptop. That sounds private, but two practical questions come first. Will the model fit in your machine's memory? Can you show that the answer was produced after the network was disconnected?

You will install a desktop application, choose an open model without relying on a soon-outdated recommendation, review its licence, download it while online, disconnect the laptop, and run one synthetic drafting task. The result may be less capable or slower than a strong cloud service. That is acceptable: the goal is a bounded local draft that you inspect yourself, not the best possible answer at any cost.

3. After this you can

• Install a desktop application for local models without using a command line.
• Choose an open model whose licence you have reviewed as compatible with this exercise and whose file fits within your laptop's available memory with room to spare.
• Run a local chat after disconnecting the laptop from every network.
• Compare the privacy, speed, quality, and convenience boundaries of local and cloud use.
• Capture one inspectable screenshot proving the bounded offline test.

4. Prerequisites

• T01-L01 · What AI can and can't do for your work.
• A Windows, macOS, or Linux laptop on which you may install approved desktop software.
• Permission to download a large model file and enough free storage for it.
• Access to the laptop's graphical system settings so you can inspect memory and disconnect Wi-Fi, Ethernet, mobile tethering, and VPN connections.
• About 30 minutes after installation and download are complete.

This exercise uses LM Studio and its graphical interface. Do not open Terminal, PowerShell, a shell, or the application's local server. Jan offers a similar desktop route and is listed in section 11, but use one application for this exercise so the test remains clear.

Use only the synthetic prompts on this page. Do not use real participant data, patient data, customer records, personnel details, credentials, unpublished findings, or confidential notes merely because the model runs locally. Your organisation's approval and data rules still apply.

5. The idea in one page

A local model is a model file processed by an application on your own computer. Once the application and model file are already present, a basic chat can run without sending the prompt to a cloud model. Change your behaviour: download everything first, then disconnect every network and test the exact local selection before considering sensitive work.

The model-selection rule has two gates. First, it must be an open model: its model weights are available to download, and its model card identifies a licence that you or your organisation has reviewed as compatible with this personal, synthetic exercise. A visible licence is not automatic permission for every use; follow its conditions, and do not use a model when the licence is missing, unclear, or not approved. Second, the model must fit in memory, with room to spare for the operating system, the desktop application, and the current conversation. A download fitting on disk does not prove it fits in working memory. Change your behaviour: check the model card and licence before downloading, check available memory rather than only storage, and begin with a smaller compatible open model file when the fit is uncertain. If loading fails, the system becomes unresponsive, or memory remains nearly full, unload it and choose a smaller file rather than repeatedly forcing the same choice.

Do not turn that rule into a universal table of laptop models or model names. Hardware, model files, formats, and software support change. A model that fits may still run slowly, and two files with similar size may behave differently on a particular system. Change your behaviour: treat the application's compatibility information as a starting point and one successful run on your exact laptop as the useful evidence. It is not evidence for someone else's machine.

Local use changes trade-offs:

A small local model can be genuinely enough when you provide all necessary text and can judge the output: rewrite a paragraph, propose headings, classify invented feedback into fixed labels, or turn synthetic notes into a draft. It is not enough merely because the material is private. Current research, reliable citations, difficult calculations, high-impact decisions, or work you cannot check need another approved process.

An offline response proves a narrow fact: that this selected local model completed this prompt without an active network connection. It does not prove that the application never connected earlier, that files and screenshots are securely stored, that other software cannot read them, or that using real data is lawful and approved. Change your behaviour: keep the exercise synthetic and treat device security, account access, backups, retention, and organisational permission as separate boundaries.

6. The worked example: one bounded offline draft

The Lab and Company versions use the same application, steps, prompt shape, and checks. Only the fictional notes and the intended reader change.

Prepare the laptop in the graphical interface

1. Open the official LM Studio website from section 11. Confirm that the site and installer publisher are the expected vendor, download the installer for your operating system, and install it through the normal graphical installer. If your organisation provides managed software, use that route instead.
2. Open the operating system's About, System Information, Task Manager, or Activity Monitor view. Note the installed memory and how much is currently available. Close games, video editors, large data tools, and unused browser windows so the model is not competing with them.
3. In LM Studio, open the model discovery view. Consider only an open model whose weights are available to download. Read its model card and identified licence, then confirm that you or your organisation has reviewed that licence as compatible with this personal, synthetic exercise. If the licence is missing, unclear, or not approved, do not download the model; choose another open model and repeat the review.
4. From the models that pass that licence check, choose one intended for conversational instructions, in a file size that fits comfortably within currently available memory and a format the application marks as compatible. Do not select by popularity, a leaderboard, or a model name copied from this course. If the file is incompatible or too large, choose a smaller open model that has also passed the licence review.
5. Download that one model while the laptop is online. Wait until the interface says the download is complete. Open the chat view and explicitly select the downloaded local model; do not select an optional cloud provider.
6. Send this harmless readiness prompt while still online:

Reply with exactly: LOCAL TEST READY

If it does not respond, fix the local loading problem before disconnecting. Do not enable a local server, API, tool, document connector, or web search. Those belong outside this Level 1 exercise.

Disconnect before the meaningful test

Use graphical system controls to turn off Wi-Fi and Bluetooth tethering, unplug Ethernet, stop any mobile hotspot connection, and disconnect the VPN. Confirm that the operating system reports no network connection. Leave that visible status available so it can appear beside the chat at the end.

Keep LM Studio open with the downloaded local model selected. Do not reload a web page or sign into anything. Choose either of the following parallel prompts.

Lab framing: unpublished-result-shaped notes

The notes are fictional. They imitate the shape of private working notes without containing any real project or person.

Use only the synthetic notes below. Draft a private working note of 80-120 words
for the fictional researcher who wrote them. Separate Observation, Possible
explanation, and Next check. Do not add causes, measurements, or conclusions.

Synthetic notes:
- Northstar bench exercise, samples L-01 to L-04
- L-03 showed a weaker signal in both repeat readings
- preparation log for L-03 has one unchecked mixing box
- no cause established
- next action under consideration: inspect the preparation record and repeat L-03

The response is useful if it preserves the uncertainty, keeps the repeated observation separate from the possible preparation issue, and does not claim that mixing caused the weak signal. The researcher checks every sentence against the synthetic notes. An example of an acceptable sentence is: Possible explanation: the unchecked mixing box is a lead to inspect, not an established cause.

Company framing: confidential-matter-shaped notes

The structure is identical: one observation, one possible explanation, and one next check. No real employee or company information is used.

Use only the synthetic notes below. Draft a private working note of 80-120 words
for the fictional manager who wrote them. Separate Observation, Possible
explanation, and Next check. Do not add motives, promises, or conclusions.

Synthetic notes:
- Harbor team exercise, cases C-01 to C-04
- C-03 remained unassigned after the second daily review
- handover checklist for C-03 has one unchecked ownership box
- no cause established
- next action under consideration: inspect the handover record and confirm an owner

The response is useful if it preserves the uncertainty, distinguishes the unassigned case from the unchecked box, and does not blame a person. The manager checks every sentence against the synthetic notes. An acceptable sentence follows the same boundary: Possible explanation: the unchecked ownership box is a lead to inspect, not an established cause.

Check the same four things

Whichever framing you used, inspect the response before keeping it:

1. Source: every factual detail appears in the supplied synthetic notes.
2. Uncertainty: the output says that no cause is established.
3. Format: the three requested headings are present and the response is within the requested length.
4. Boundary: the selected model is local and the operating system still shows no network connection.

If the answer fails a content check, revise the prompt or choose a smaller, clearer task. Do not reconnect merely to obtain a stronger cloud answer and then describe that output as local. When the check is complete, stop or unload the model so it does not continue using memory, battery, or accelerator resources.

7. What goes wrong

The model file is too large

Symptom: loading fails, the laptop freezes, other applications become unusable, or the system starts moving data heavily between memory and disk.

Fix: unload the model, close unnecessary applications, and choose a smaller compatible model file. Fitting on disk is not the sizing test; available memory with headroom is.

Frontier quality is expected from a small local model

Symptom: a weak answer to a difficult research, legal, numerical, or strategic question is treated as proof that local models are useless.

Fix: test a bounded transformation such as drafting, rephrasing, or fixed-label classification. Use an approved stronger process when the task exceeds what you can verify locally.

Local is mistaken for automatically permitted

Symptom: real colleague, participant, customer, or unpublished data is pasted because no cloud model was selected.

Fix: apply the same paste rules as before. Confirm purpose, permission, device security, storage, backups, and deletion. Until then, use public, synthetic, or explicitly approved data.

A connected feature silently changes the route

Symptom: the selected chat uses a cloud provider, web search, document connector, or remote service instead of only the downloaded model.

Fix: select the local model explicitly, leave integrations disabled, disconnect every network, and confirm that the response still completes. One local-looking interface can contain several different routes.

Heat and battery are ignored

Symptom: the laptop becomes hot, fans stay loud, or the battery drains after the conversation ends.

Fix: place the laptop on a ventilated surface, plug it into approved power when appropriate, stop if the operating system reports a thermal problem, and unload or stop the model after use.

The proof exposes the thing you meant to protect

Symptom: the screenshot contains real sensitive text, a username, notifications, file paths, or another application's content.

Fix: use only the synthetic prompt, close unrelated windows and notifications, and inspect the image before retaining it. Evidence should prove the test without creating a new disclosure.

8. Do it yourself: a 30-minute offline run

Complete this only after the desktop application and one suitably sized open model with a reviewed, compatible licence are downloaded.

Minutes 0-5: confirm that the downloaded model's weights were openly available and that its identified licence was reviewed as compatible with this exercise. Then close heavy applications and check currently available memory, model file size, free storage, and the application's compatibility indication. If either the licence or technical fit is doubtful, stop and choose another open model that passes both checks before continuing.

Minutes 5-9: open a new chat, explicitly select the downloaded local model, and send Reply with exactly: LOCAL TEST READY. Confirm that no cloud provider, web search, connector, local server, or tool is enabled.

Minutes 9-13: turn off Wi-Fi and tethering, unplug Ethernet, disconnect the VPN, and confirm the operating system reports no network connection.

Minutes 13-21: run either the Lab or Company prompt from section 6. Keep the material synthetic. Wait for the complete response without reconnecting.

Minutes 21-26: check source fidelity, uncertainty, format, and the selected local route. If a check fails, make one prompt correction and retry while still disconnected.

Minutes 26-30: arrange the local chat and graphical network status so both are visible at once. Close notifications and unrelated applications, take the required screenshot, inspect it for accidental disclosure, then stop or unload the model.

9. Exit check

Deliver exactly one artifact: one evidence screenshot showing all of the following in a single image:

• the LM Studio chat with the downloaded local model visibly selected;
• the complete answer to either synthetic prompt from section 6;
• the operating system's graphical status showing that no network is connected; and
• no personal, confidential, unpublished, customer, participant, patient, or credential data.

It passes when another person can identify the local application and selected model, read enough of the synthetic answer to see that generation completed, and see the disconnected state in the same image. The screenshot proves this one offline run only; it is not a general privacy certification.

10. Rule to remember

Fits in memory, or it does not run.

11. Further reading & tools

• Taught: LM Studio desktop documentation (opens in a new tab) - primary vendor guidance for installing the desktop application, downloading a model, checking current system requirements, and operating offline; verify the current interface before use.
• Taught: Hugging Face model cards (opens in a new tab) - primary platform guidance for reading a model's intended uses, limitations, licence information, and other metadata before download.
• Catalogued: Jan source repository and documentation (opens in a new tab) - primary project source for the alternative offline-first desktop application; follow its current graphical setup rather than assuming LM Studio's controls are identical.
• Taught: Local & Private Models - separates a local model from the wider device, network, identity, retrieval, logging, and operator boundaries.
• Taught: T01-L01 · What AI can and can't do for your work - chooses a bounded task and keeps output checking with the user.
• Taught: T12-L01 · What you may and may not paste into AI at work - applies privacy and security boundaries before any real material is used.
• Catalogued: T11-L02 · A local model your tools can call - continues to APIs and tool connections after this GUI-only personal run.
• Catalogued: Tools index - compare current local desktop options only after checking approval, hardware fit, data handling, and primary documentation.