Run a local model and check the data route
Can I run this task on my computer, and what still needs protection?
Download a suitable model, choose local inference, and test a fresh synthetic task with the network disconnected. Then check saved files, integrations, and permission to process the material.
In brief
- Downloading a model and running it are separate operations.
- A local interface can still use cloud models, web search, or connected tools.
- Local processing does not grant rights to the input or remove device and storage risks.
Decide whether local execution fits
A local model can be useful for a small repeatable task, an offline exercise, or learning how model choice affects an answer. You take on installation, downloads, storage, and updates. Choose a task whose result you can check, such as rewriting a paragraph you wrote or explaining a calculation with a known answer.
Hardware support depends on the application, model, context size, and operating system. Consult the current LM Studio requirements or Ollama hardware documentation before installing. A supported computer does not guarantee that every model fits.
Scroll horizontally to view all columns.
| Choice | Practical consequence | Check |
|---|---|---|
| Model size | The model needs disk space and working memory. | Download size, available memory, and the task you want it to perform. |
| Quantization | Compressed model variants reduce file size with a possible quality trade-off. | Exact variant and whether its answers survive your task checks. |
| Context size | A longer conversation or document can require more memory. | Start with a short prompt and modest context; increase only when needed. |
| CPU and GPU | The execution route affects available memory and response speed. | Supported hardware and the runtime actually selected. |
| Model licence | Download availability does not settle permitted reuse. | The model card and licence for your intended use. |
The quantization explanation follows LM Studio's download documentation. Ollama's quickstart describes memory and CPU fallback for its current example model. These are vendor instructions, not an independent comparison of model quality. Evaluate the exact model on your task.
Try LM Studio
Use synthetic text for your first run. Do not begin with a real submission, interview, internal document, or lecture recording.
- Install LM Studio from its official website and check the linked system requirements.
- Open Discover. Search for a supported model by name or use the model's repository URL. Read its model card, licence, download size, and available variants.
- Download a variant that fits your computer. Obtain the required runtime while you are connected. Keep a note of the application version, exact model, variant, and context setting.
- Select the downloaded model in a chat and load it. Leave connected tools disabled for this first exercise.
- Ask a short question using material you created. Check the answer against something you already know.
These steps follow the vendor's download guide and offline documentation. Searching, downloading models or runtimes, and update checks use a network. Downloaded-model chat can work offline according to those documents.
LM Studio also supports MCP integrations. Some can execute code, read local files, or use the network. Review each integration and its permissions before enabling it. The core chat's documented offline behaviour does not describe every connected tool.
Try Ollama as an alternative
Install Ollama from its official download page, then follow its quickstart. Select a local model. The product also supports cloud models, so confirm the selected route.
For the terminal workflow, replace the placeholder below with a local model tag from the model library:
ollama run <local-model-tag>
The first run can download the model. Complete that download while connected, then use the same model for the offline exercise below. Check the model's licence and hardware needs before choosing it.
Ollama documents a local-only setting in its FAQ. Add the following to ~/.ollama/server.json, preserving any other settings, and restart Ollama:
{
"disable_ollama_cloud": true
}
The FAQ also describes OLLAMA_NO_CLOUD=1 as an alternative. These settings disable Ollama cloud models and web search. They do not control other applications you connect to Ollama.
Keep the server bound to its documented default, 127.0.0.1:11434, for this exercise. Changing OLLAMA_HOST, adding a proxy, or using a tunnel can make the server reachable elsewhere. You can inspect loaded models with:
ollama ps
The Processor column reports CPU, GPU, or mixed loading. It helps explain the execution route; it does not measure answer quality.
Verify a fresh task offline
- Finish downloading and loading the model. Record its name and settings.
- Disable cloud features and connected tools. Start a fresh chat.
- Disconnect Wi-Fi and any wired network connection.
- Send the sample note. Confirm that the answer completes and compare it with the original.
- If you need document chat, repeat with a small file containing only this synthetic note.
- Record what worked, what failed, and which functions were enabled. Close the exercise before reconnecting.
A successful result shows that this demonstrated task can complete without a live network connection. It does not audit all application traffic, connected software, or later behaviour after reconnection. A failed result may mean that a model, runtime, or tool still needs a remote service. Investigate the selected route before using real material.
Check the whole processing route
Use the same questions for a local application, an institutional service, and a consumer account. Complete them for each enabled stage, including document processing, retrieval, tool calls, and logging. This table is an editorial decision aid.
Scroll horizontally to view all columns.
| Question | Evidence to inspect |
|---|---|
| Where is data processed and stored? | Inference location, storage location, and each connected service. |
| What records are kept? | Chat history, documents, derived indexes, exports, logs, and usage records. |
| Who can access them? | Device users, service operators, support roles, and other recipients. |
| Are inputs used for training? | Terms for this product and account, including feedback and optional features. |
| How long are records retained? | Separate periods for each category, including backups. |
| What does deletion remove? | Main records, derived stores, exports, synced copies, and backup treatment. |
For local work, inspect saved chats and exports, shared computer accounts, device access, and cloud sync or backup settings. A local calculation can still create a copy in a synced folder. Deleting a chat does not establish that every exported copy has disappeared.
Permission still applies
TUM ProLehre's current guidance says students need instructor consent before processing course materials with AI and should not enter personal, sensitive, or confidential information, including into approved tools. Local execution does not create an exception to that guidance.
For research material, ask the project owner about confidentiality, participant information, and agreed uses. The TUM Research Data Hub links advice on legal aspects, licences, and anonymisation. Removing names alone does not establish that an interview is anonymous.
Use the rules and rights guide to check inputs and reuse. If the permission or route is unclear, continue the exercise with synthetic material while you ask the responsible person.
Sources and review date
Vendor pages were checked on 5 October 2026. They document supported functions and settings; this page does not report a tested TUM installation or certify a privacy configuration. Interface labels, model tags, hardware support, and defaults can change. Recheck the linked instructions when you install or update.
The offline exercise and six-question table are editorial examples. TUM guidance supplies the institutional boundary; it does not approve LM Studio, Ollama, or a particular model for protected data.