Local AI for confidential files: run a model on your office PC or Mac with LM Studio or Ollama (lesson 9)

Yes. A free program such as LM Studio or Ollama can run an AI model on an office computer itself, so what you type and the files you attach stay there instead of travelling to an AI company. You need a fairly recent PC or Mac; LM Studio recommends 16 GB of memory. The catch: a model small enough for an office computer makes more mistakes than the big online assistants, and on an ordinary computer it can be slow. Both programs also sell online extras that send your text out; keep them off. Everything below was checked on the makers’ pages on 3 November 2026.

This site is written with AI tools, including Claude by Anthropic. This page uses only what makers, public authorities and published studies say, with links.

This is lesson 9 of the free course “AI for small business”. Lesson 8 was about agents, AI that acts for you on a computer; this lesson is for offices whose files must stay confidential, such as a law firm, an accountant or a medical practice.

What does “running AI locally” mean?

It means the language model, the program that writes the answers, works on your own computer instead of a company’s servers. With an online assistant such as ChatGPT, Claude or Gemini, your text travels to the maker’s computers and the answer comes back. With local AI, the model sits on your own disk; the internet is needed only to download it and the program that runs it.

LM Studio’s offline page says that with a downloaded model “Nothing you enter into LM Studio when chatting with LLMs leaves your device”, documents included. Ollama’s FAQ says: “We don’t see your prompts or data when you run locally.”

For professionals, this answers the worry in lesson 2: the Council of Bars and Law Societies of Europe warns that lawyers may breach their duties “already at the point of inputting the data into a system”. On your own computer, no outside company receives the data. The other rules stay: the privacy law still applies, and in Italy Law 132 of 2025 still asks professionals to tell clients which AI systems they use.

What is an open-weight model, and may your office use it?

A language model is a very large file of numbers, adjusted while the maker trains it on huge amounts of text. These numbers are called weights: together they hold what the model learned, like the many knobs of a mixing desk that together make the sound. An open-weight model is one whose maker publishes that file, so anyone can download it.

Each model comes with a licence that says what you may do with it. On 3 November 2026 the makers’ pages said, for five well-known families:

Model (maker)	Licence	Work use
Gemma 4 (Google)	Apache 2.0; older versions use the Gemma Terms of Use	Allowed
gpt-oss (OpenAI)	Apache 2.0, plus a usage policy asking you to “comply with all applicable law”	Allowed
Llama 4 (Meta)	Llama 4 Community License and Acceptable Use Policy	Not for people or companies based in the EU; elsewhere with conditions, see below
Mistral Small 4, Ministral 3 (Mistral AI)	Apache 2.0; Mistral Large 4 has its own licence	Allowed for these two
Qwen3 (Alibaba Cloud)	Apache 2.0	Allowed

The Apache License 2.0 lets you use the model “no-charge, royalty-free”, at work too.

Llama 4 asks more. Meta’s use policy says that for multimodal models, which also read images, the licence rights “are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union”, except for “end users of a product or service that incorporates” them. Meta’s model card says the Llama 4 models “are natively multimodal”, so the simple choice for an office in the EU is another family. Elsewhere, whoever passes the model on must show “Built with Llama”. A licence covers the model, never your data.

LM Studio or Ollama: which program?

Both run models on your computer for free, in a window with a chat like an online assistant. From their own pages:

	LM Studio	Ollama
Cost	Free “both at home and at work” since 8 July 2025 (LM Studio); terms for “personal and / or internal business purposes”	Free, open-source (MIT licence); running models on your own computer “is always unlimited” (pricing)
Paid online extras	Bionic+ 20 dollars a month, Pro 100, for models hosted in the US and web search	Pro 20 dollars a month, Max 100, for cloud models
Windows	Intel or AMD processor with the AVX2 feature (ask the PC’s maker), or Snapdragon X Elite (requirements)	Windows 10 22H2 or newer (Ollama)
Mac	Apple silicon, macOS 14 or newer; not Intel Macs	macOS 14 or newer; Apple silicon, or Intel using the main processor only (Ollama)
Keep off for confidential work	Cloud models, web search	Models with “cloud” in the name, web search

Neither pricing page says whether tax is included.

How much memory does your computer need?

Memory, or RAM, is the computer’s working space while programs are open, not the disk where files are kept. The model must fit in it, with room left for your document and other programs. To check: on a Mac, Apple menu, System Settings, General, About, then “Memory”, and “Chip” for Apple silicon or “Processor” for Intel (Apple); on Windows, right-click Start, Settings, System, About, then “Device specifications” (Microsoft).

Models come in compressed copies. LM Studio’s download guide says a Q in the name, such as “Q_8”, means quantization, “compressing model files in size, while giving up some degree of quality”, and advises a 4-bit option or higher.

Source	What it says
LM Studio, Mac	16 GB or more recommended; it “may still” work on 8 GB Macs if you “stick to smaller models and modest context sizes”
LM Studio, Windows	At least 16 GB recommended, and a graphics card with at least 4 GB of its own memory
OpenAI	gpt-oss-20b can “run within 16GB of memory”
Google	To load Gemma 4 in 4-bit on a graphics chip: E2B 2.9 GB, E4B 4.5 GB, 12B 6.7 GB, 26B A4B 14.4 GB, 31B 17.5 GB

Google’s figures cover only the model itself, not other programs and not the text being read, which needs “significantly more” memory as it grows. So choose a model whose figure is well below your computer’s memory.

How do you start with LM Studio?

From LM Studio’s getting started page:

Install it from lmstudio.ai/download. Choose “Download LM Studio”, not “LM Studio Bionic”, which the page calls “An agent made for open models”.
Download a model in the Discover tab: search a name such as “gemma” and pick a Q4 or higher option that fits your memory.
Load it in the Chat tab: open the model loader (Cmd+L on a Mac, Ctrl+L on Windows) and select the model. Before loading you can change the load settings, including the context size (LM Studio), how much text the model reads at once; more uses more memory.
Add a document: drag a .docx, .pdf or .txt file into the chat (Chat with Documents).
Test it offline: switch off Wi-Fi. The offline page says chats and documents work without the internet.

If a document is too long to read at once, LM Studio picks only the passages that look relevant, so a summary can miss parts. Its privacy policy says that with local models “none of your messages, chat histories, and documents are ever transmitted from your system”; the app still sends details such as its version and your internet address when it downloads models or checks for updates. Leave its cloud models and web search off.

How do you start with Ollama?

From Ollama’s quickstart and its app announcement:

Install it from ollama.com/download: on a Mac open the downloaded file and drag Ollama into the Applications folder; on Windows run OllamaSetup.exe.
Choose a local model. The setup says: “Sign in to use cloud models, or choose a local model.” Choose a local one; you do not need to sign in.
Avoid models with “cloud” in the name, such as gemma4:cloud (Ollama): they run on Ollama’s servers, mainly in the United States, its pricing page says.
Make room for long documents: “Change the slider in the Ollama app under settings” (context length). With less than 24 GB of graphics memory, Ollama starts at “4k” tokens, pieces of words: roughly 3,000 words of English text, as LM Studio counts a token as about three-quarters of a word.
Add a document: the app supports drag and drop of text files and PDFs.
Lock out the cloud: ask whoever looks after your computers to switch on the “local only mode” in the FAQ, which turns off cloud models and web search.

Example: how do you summarize a 30-page contract in the office?

A law office receives a 30-page commercial lease a client wants to sign, and the lawyer wants a one-page summary before the meeting.

Load the largest model your memory allows, and raise the context length (the load settings in LM Studio, the slider in Ollama): a 30-page contract is usually far longer than 3,000 words.
Drag the PDF into the chat and send:

Read the attached contract. Write a summary of at most one page, in plain [English], under these headings: parties; what the contract is about; duration and renewal; prices, payments and due dates; what each party must do; penalties; how the contract can end; liability and guarantees; which court or arbitration. After each point, give the clause number. If a heading is not covered, write “not found”. Copy amounts and dates exactly as written. Do not say whether the clauses are fair or legal.

Then ask:

List every date, deadline and notice period in the contract, in date order, with the clause number.

Check: open every clause cited, and look for the “not found” headings yourself. The legal assessment stays yours (lesson 2).

Example: how do you anonymize a document before using an online assistant?

A medical practice wants an online assistant to rewrite a long letter about a patient in simpler words. The local model removes the personal details first.

In LM Studio or Ollama, paste the letter with:

Replace every personal detail in the text below with a placeholder, and change nothing else. Use [PERSON 1], [PERSON 2] for people, [COMPANY 1] for companies, and [ADDRESS], [PHONE], [EMAIL], [TAX CODE], [DATE OF BIRTH] for the rest. Use the same placeholder each time the same person appears. After the text, list each placeholder with the original value. Text: [paste]

Read the result line by line. A local model can miss a name or a number.
Look for details that point to one person: a rare job, a small village, an exact date, a rare illness. Personal data covers anyone “identified or identifiable” (EDPB), even without a name. Make such details general, or keep the job local.
Paste only the text, never the list, into the online assistant, ideally on a business plan (lesson 3).
Put the real details back by hand on your computer.

What can go wrong with local AI?

More mistakes. Google says larger, less compressed models “are generally more capable” (Gemma); the ones that fit an office computer are small. They can also invent facts, like every chatbot (lesson 1).
Slower answers. For best performance, Ollama says to “avoid offloading the model to CPU”, the main processor (Ollama): the model should run on a graphics chip with enough memory of its own. On a computer without one, answers come more slowly.
No web search. A local model knows only what it learned and what you give it; in both programs web search is an online feature.
A stolen laptop. Local AI keeps data away from AI companies, not from a thief: chats are “saved locally on your device”, LM Studio says.

So protect the computer:

Turn on disk encryption, which makes the disk unreadable without your password. Mac: Apple menu, System Settings, Privacy & Security, FileVault; keep the recovery key safe, because without password and key the files cannot be recovered (Apple). Windows, signed in as an administrator: Settings, Privacy & security, Device encryption; if it does not appear, ask whoever looks after your computers (Microsoft).
Keep encrypted backups and install updates of the system and the AI program.
Lock the screen when you leave, and delete chats you no longer need.
Do not open the AI to the office network without a technician; Ollama answers only its own computer unless a setting is changed.

When is an online business plan the better choice?

Choose an online assistant on a business plan, which comes with a data protection contract and can keep your work out of AI training (lesson 3), when:

the job needs the most capable models, such as long reasoning across many documents;
you need recent information from the web;
several people need AI on phones and laptops;
your computers are old or short of memory, or nobody can keep them encrypted and updated.

The two can also work together, as in the second example.

Where to go next

Lesson 2: the rules explains privacy, professional secrecy and Law 132.
Lesson 3: plans and costs compares business plans and their contracts.
Running a local LLM on a Raspberry Pi goes deeper into memory and compression.
AI for everyone is the complete guide to today’s AI.

Quiz: check what you learned

Pick one answer for each question: you see at once whether it is right, and why. Your answers are not sent anywhere.

In the Ollama app, which model runs on Ollama's servers instead of your computer?
Show the answer
Right answer:B. gemma4:cloud
Ollama's cloud page names gemma4:cloud as the cloud model to use in the app. Cloud models run on Ollama's servers, mainly in the United States.
What does local AI NOT protect against?
Show the answer
Right answer:B. Someone reading the chats on a stolen laptop without disk encryption
Chats stay on the computer. If it is stolen and the disk is not encrypted, they can be read.
A local model has replaced the names in a letter with placeholders. What do you do next?
Show the answer
Right answer:A. Read the whole text for anything it missed, then paste only the text, not the list of original values, into the online assistant
A local model can miss a name or a detail that identifies someone. The list stays on your computer; you put the real details back by hand.
What does Meta's Llama 4 use policy say about multimodal models, the ones that also read images?
Show the answer
Right answer:B. The licence rights are not granted to people domiciled in the EU or companies based there
The policy says the rights are not granted to an individual domiciled in, or a company with its principal place of business in, the European Union. The only exception is for end users of a product or service that incorporates those models.

Common questions

Is LM Studio free for business use?

Yes, for models that run on your own computer. LM Studio announced on 8 July 2025 that the app is free "both at home and at work", and its terms of 23 August 2026 cover "internal business purposes". Only its online extras are paid.

How much RAM do I need to run AI locally?

LM Studio recommends 16 GB, and says 8 GB Macs may still work with smaller models. Google gives 2.9 GB to 17.5 GB for its compressed 4-bit Gemma 4 models, for the model alone; other programs and the text being read need more.

Does Ollama send my data to the internet?

Not with a model downloaded to your computer: Ollama's FAQ says "We don't see your prompts or data when you run locally." Models with "cloud" in their name, and web search, run on Ollama's servers. A local-only mode, described in the FAQ, switches those features off.

Topics: AI for small business, local AI, LM Studio, Ollama, open-weight models, privacy, professional secrecy, GDPR