Run LLM locally when the text should stay on a machine you control, and when you are willing to read what that machine actually does. A local app is not automatically private if it phones home or downloads a cloud model. The three runners to know are Ollama, LM Studio, and llama.cpp. The model card still lives on a hub, not inside the chat window. Choosing the model itself is Best Open-Source LLMs for Practical Work.
Key takeaway: Pick a runner you can inspect, read the model card before the download, and confirm the app is not calling a cloud model. On a Mac, keep the existing AI Savr M4 guide for that machine. This page is the PC runtime.
Research reviewed: October 30, 2026. This is desk research by the AI Savr editorial team. We did not run a private lab test, and we did not invent scores, traffic numbers, or hands-on results. Confirm live prices on the vendor page before you buy.
What does running a model on your PC actually mean?
It means the weights you selected are executed on that computer, and the prompt is not sent to a vendor API as the main path. It does not mean every feature in the app is offline. Some apps check for updates, log in, or offer a hosted model beside the local one. If you click the hosted model, you are not running locally anymore. Read the card for license and size before you download, using the steps in Best Open-Source LLMs for Practical Work.
| Need | Open | Privacy check |
|---|---|---|
| A simple local chat | Ollama | Confirm the model is local, not cloud |
| A desktop app | LM Studio | Read what the app may send |
| The model file itself | llama.cpp | You choose the file you run |
| The card and the download | Hugging Face | The hub is not the runtime |
Runners and the place you read the card
Ollama for a straightforward local chat
Ollama is a common way to pull a model and chat with it on your own computer. That convenience is the point, and it is also the risk. Read the screen before you accept a download. If the choice is a cloud model, you are not staying local, even though the app is installed on the PC.
Choose it when: you want a local chat or a local API for a tool you already run.
Leave it when: you have not read the model card, or the data cannot leave the machine and you have not checked the app's network use.
Price and access: Confirm the license, the price, and any cloud add-on on ollama.com before you install it on a work PC. A free app can still offer a hosted model.
A real use: Install it on a test user account, pull one small model the card allows, and summarize a paragraph you wrote. Watch whether the app offers a cloud model in the same menu. Do not point client text at that option.
LM Studio for a desktop window
LM Studio is a desktop app for people who want a window, a model browser, and a local chat without living in a terminal. A window on your PC is not a privacy proof. Check the settings for sign-in, telemetry, and any feature that sends the prompt out. The vendor page is the source for what the build does.
Choose it when: the person running it will not use a command line.
Leave it when: you need a reproducible server and a named model file more than a desktop browser.
Price and access: Confirm price and privacy notes on lmstudio.ai before you install. Do not assume a desktop app is offline.
A real use: Load one model you already chose from its card. Ask a question you know the answer to, from a paragraph on disk. Then disconnect the network and see which features still answer. If the answer requires the network, it was not the local path.
llama.cpp when you want the file and the hardware
llama.cpp is the project to open when you want the model file, the build, and the hardware limits in front of you. It does not pick a winning model. You bring a file that fits the machine. If the model is too large, the run fails or becomes unusable. That is a clearer outcome than a quiet switch to a cloud model.
Choose it when: someone on the team can follow a repository and you care which file is loaded.
Leave it when: nobody can maintain a build. Use a desktop app and still do the privacy check.
Price and access: Read the LICENSE file in the repository before you ship it inside a product. Do not assume a GitHub project is free for every commercial use without reading that file.
A real use: Pick a small model the card says will fit your RAM. Run a prompt you wrote. If the process is swapping until the machine stalls, the model is too big. Choose a smaller file. Do not 'fix' it by sending the prompt to an API.
Hugging Face for the card, not the runtime
Hugging Face is where you read the model card and download weights. It is not a local runner. The card is how you check license, size, and whether the file is the model you think it is. A hub account can also gate a download. Accepting that gate is not the same as running the model offline.
Choose it when: you need the card, the license, and the file name before you touch Ollama, LM Studio, or llama.cpp.
Leave it when: you want a chat window. Go back to a runner after the card is read.
Price and access: Reading cards is free. Some files ask you to accept a license before download. Confirm that step on the card. This guide does not quote a paid hub plan.
A real use: Open the card, write down the license and the file size, and only then download. If the card says the weights are not for your use, stop. Do not look for a mirror that skips the card.
Install once, without fooling yourself
A local setup is a file, a runner, and a network check. Skip any one of those and you may be in the cloud with a desktop icon.
- Read the card first. License, size, and the exact file name. The card beats a roundup, including this one.
- Match the file to the PC. If it does not fit in memory, pick a smaller model. Do not start and hope.
- Choose one runner. Ollama, LM Studio, or llama.cpp. Learn one before you install three.
- Test with the network off. After the model is on disk, a local answer should still work offline. If it does not, you are not on the local path.
- Keep secrets out of the prompt anyway. A local model can still be copied, logged by the app, or pasted into a ticket. Local is not a vault.
Local is a place, not a promise
Running on your PC does not make the app private, and it does not make the model correct. If the app phones home, or if you download a cloud model, the prompt can leave the machine. Model names change quickly. The companion for an M4 Mac is the existing AI Savr guide named in the research note for this article. This page does not repeat benchmark scores from that guide or from anywhere else. Read the card for the build you actually downloaded.
What a team with sensitive notes should do
Pick one PC, one runner, and one small model that the card allows for the work. Write a one-page rule: which prompts may be local, which features must stay offline, and who may update the app. Do not point a local demo at customer exports on day one. If the work later becomes an agent that can send mail or merge code, read AI Agent Security for Safe Automation before you connect anything.
Questions worth asking before you buy
Ask whether the build can run offline, whether there is a cloud model in the same menu, and what the license allows if you ship a product. Ask who on your team can update the app when the project ships a breaking change.
For companies that need more than a subscription
If several people need the same local model, standardise the file name and the runner. A custom agent can wrap an approved local model with a log. It should not add a cloud fallback in secret.
If an off-the-shelf tool cannot meet your data, approval, or security rules, talk to AI Savr about a custom AI agent. Use the directory to compare live tools first, and vendors can submit a product.
Frequently asked questions
What does it mean to run LLM locally?
The model file you chose runs on your PC. The prompt is not sent to a vendor API as the main path. You still have to check that the app is not using a cloud model.
Is local automatically private?
No. Local is not private if the app phones home or if you download a cloud model. Test the feature you care about with the network off.
Which app should I install first?
If you want a simple pull-and-chat flow, start with Ollama. If you want a desktop window, start with LM Studio. If you want the file and the build, start with llama.cpp.
Where do I read what the model allows?
On the model card, often on Hugging Face, before you download. The runner does not replace the card.
Can a local app still call the cloud?
Yes. Many apps offer a hosted model beside the local one. If you select it, the work is not local.
What about a Mac?
Use the existing AI Savr guide for local models on an M4 Mac for that machine. This page is about PC runners, not a benchmark table.