Your first local AI in 15 minutes

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

A large language model (LLM) predicts the next word from your prompt and keeps writing from there. Ollama is a free program that downloads these models and runs them on your machine. Leave the service off the public internet, and the conversation stays in the house. The files are often GGUF, a compressed form (quantization such as Q4 or Q5) that takes less disk space and less RAM. VRAM is the memory on the graphics card. When part of the model fits there, replies get much smoother.

This walkthrough aims for a concrete result in 12 to 17 minutes, including the ~2 GB model download: Ollama installed, a small model pulled from the official library, and a first exchange in the terminal. The “15 minutes” in the title is the typical time on fiber. Plan on about 17 minutes if the download is slower. The steps below are written for Windows and PowerShell. The standard Ollama install does not ask for administrator rights. On Mac or Linux, the same ollama commands work after you install from ollama.com/download. The multi-OS guide covers those variants.

For a more detailed Windows install (GPU drivers, where the files land), continue with install Ollama on Windows.

Which page should you read?

Goal Page Rough time
A first timed success on Windows This guide 12–17 min
Windows, step by step (GPU, PATH, fixes) Install Ollama on Windows ~25 min
Mac, Linux, or more than one OS Ollama on every OS ~20–30 min

Hardware during the timer: you can finish this on the PC you already have. Hardware links and partner links come after the recap, once the Ollama test has worked.

Who this is for

You What you get What you need
Curious, and not ready to buy hardware Proof that local AI runs on the living-room PC 64-bit Windows, 16 GB RAM, internet
A household that wants a nicer screen later A first success in the terminal, before Open WebUI The same base, plus about 20 minutes free
Already on a Mac or Linux The same ollama commands after install. This page spells out Windows See Ollama on every OS

Checkpoint table (end of the walkthrough)

Use this table at the end of step 4 so you can confirm the timer without rereading the whole page.

Check What you want to see
Ollama is running ollama –version prints a version number in the console
Model downloaded The tag llama3.2:3b shows up in ollama list
First chat At least one question and answer through ollama run
Quality of the reply No “out of memory” error. The pasted prompt asks for French: at least three coherent French sentences that mention running locally and a privacy or data benefit. Hallucinations can still happen. This test checks the install, not legal or medical facts.
Total time 12–17 min depending on download speed (model download ~2 GB)

Minimum requirements (no new hardware)

  • Windows 10 or 11, 64-bit. The standard Ollama install, in your user folder, does not ask for administrator rights.
  • 16 GB of RAM or more, and 20 GB free on a fast drive (an SSD is the comfortable choice).
  • Internet for the one-time download of Ollama and the model.
  • A company antivirus or firewall: plan on an exception if the installer is blocked.

If you are unsure which model size to try after this test, choose a first model lines up RAM, VRAM, and what you actually want to do.

The whole timer

Phase Rough time Running total
PowerShell and checks 2 min 2 min
Ollama installer 3 min 5 min
Model download (step 3) 5–10 min 10–15 min
First chat (step 4) 2 min 12–17 min
Reading the fixes (only if you need them) +5 min outside the timer

Start a timer if you want the challenge. The slow part is almost always the model download.

Mistakes that push you past 15 minutes

Mistake What happens Fix
Copy ollama -- version with a space Command not found Use the text block in step 2: ollama --version
A model tag with spaces (llama3. 2 :3b) pull fails Copy from the text blocks in steps 3 and 4. The tag is continuous: llama3.2:3b
Ollama closed during the pull connection refused Start Ollama from the Start menu and wait for the taskbar icon
The console was opened before install and never reopened PATH was not refreshed Close PowerShell and open a new window (a normal session is enough) until ollama --version answers

Step 1 — PowerShell and checks (~2 min)

  1. Right-click the Start button, then Terminal or Windows PowerShell (a normal session, not “Admin”).
  2. Check the architecture:
[Environment]::Is64BitOperatingSystem

The output should be True. If you get False, this 64-bit Windows guide does not apply as written.

  1. Optional: confirm that enough RAM is visible:
(Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory / 1GB

A value of 15 or more is comfortable for a 3B. Under 12, pick an even smaller model after you read 32 vs 64 GB of RAM for LLMs if you are thinking about an upgrade.

Do you need an administrator account? The standard install does not. The official Ollama documentation for Windows says the OllamaSetup.exe installer lands in your user profile without administrator rights. A normal PowerShell window is enough for this guide. Running Ollama as a Windows service (for example from the standalone archive with a tool such as NSSM) is a separate setup, useful for advanced use, and that one can ask for administrator rights. It is not the path described here.

Step 2 — Install Ollama (~3 min)

  1. In the browser, open ollama.com/download. The official Ollama documentation is there too.
  2. Download Ollama for Windows and run the .exe.
  3. Follow the wizard. The default location is the right choice.
  4. At the end, look for the Ollama icon near the clock. If it is missing, open Ollama from the Start menu.
  5. Close the PowerShell window and open a new one (a normal session) so the ollama command is recognized. That refreshes PATH.

Quick test. Copy the whole line, with no space inside the flag:

ollama --version

A version number should appear. If it does not, restart the PC once before you dig further.

Standard install and Windows service. The usual install (step 2) runs without administrator rights, in your user folder, as the official docs describe. Once ollama --version answers in a new console (PATH refreshed), steps 3 and 4 continue in the same ordinary Windows Terminal. The Ollama app (the icon near the clock) needs to stay running. Installing Ollama as a Windows service is a separate setup, and it can require administrator rights. It sits outside this 15-minute path.

Step 3 — Download the model (~5–10 min)

Open PowerShell (a normal session, as in the note above) and run:

ollama pull llama3.2:3b

Paste one whole line. The tag is llama3.2:3b (a dot between 3 and 2, a colon before 3b, no spaces).

You will see a progress bar for each layer. At 100%, list the models:

ollama list

The NAME column should contain llama3.2:3b. This model is a teaching compromise: light enough for a first success, capable enough for everyday language. To compare other sizes later, see VRAM and model size.

Storage: the weights pile up on the system drive. If you are short on space, read SSD storage for LLM models before you stack tens of gigabytes.

Step 4 — First chat in the terminal (~2 min)

Start it:

ollama run llama3.2:3b

Same tag as step 3. Copy it from the text block above, not from a screenshot.

After it loads (10–40 seconds depending on CPU or GPU), the >>> prompt appears. The block below asks for French on purpose, so you can see that the model followed the language you requested. Once that reply looks right, ask the same thing in English.

Test prompt (copy and paste)

Tu es un assistant concis. En trois phrases maximum, explique ce qu'est une IA locale chez soi et cite un avantage pour la vie privée. Réponds en français.

The shape of a good reply (not word for word)

Check the Quality of the reply row in the checkpoint table. A short example of the shape:

Une IA locale tourne sur votre PC : vos échanges restent chez vous si vous n’exposez pas le service.
Avantage typique : plus de contrôle sur la confidentialité des brouillons ou des questions familiales.

To compare other model sizes after this test, see choose a first model.

To leave the interactive session:

/bye

Step 5 — See whether the GPU is helping (optional, ~1 min)

Open Task Manager, then the Performance tab, then GPU. Run ollama run again with the same tag as step 4 and ask a long question. On a PC with NVIDIA and recent drivers, GPU use can climb. On CPU alone, only processor load rises. The guide still works. It is simply slower.

A web page for the household (outside the 15 minutes)

If you would rather chat in the browser, follow Open WebUI for beginners once Ollama is stable. If you are comfortable with containers, you can also stack Docker, Ollama, and Open WebUI.

To try models with a graphical app and no command line, install LM Studio sits alongside this path. Ollama fits home APIs and Home Assistant + Ollama better. LM Studio fits a visual tryout of GGUF files.

Quick fixes

Symptom What to try
ollama not found Open a new PowerShell window (a normal session) or restart. Confirm the Ollama app is running (icon near the clock).
ollama --version fails Start Ollama from the Start menu, then retry with the text block above.
connection refused on pull Start Ollama from the Start menu.
Download is very slow Fiber is busy, or Wi-Fi is weak. Let it finish, or try again later.
Reply is extremely slow Normal for a 3B on CPU. Close heavy apps.
Out-of-memory error The model is too big for the RAM. Stay on 3B, or see choose a first model.
Antivirus blocks the exe A temporary exception on the Ollama install folder.

Leave the Ollama port closed on your router until the service has a login. For a calm household setup, read privacy when AI stays home before you drop sensitive documents into a web interface.

Past the 15 minutes?

You just proved that local AI is something a household can run. Sensible next steps:

  1. Tune the model (choose a first model).
  2. Make it nicer to use (Open WebUI).
  3. A real job for it (a private family assistant, Whisper transcription).

Safety and habits

  • Keep Windows and the NVIDIA, AMD, or Intel GPU drivers updated.
  • Keep your Ollama instance off the public internet until it has a password and HTTPS.
  • Treat model replies as drafts. Check the facts for health, money, or paperwork.

Recap

In 12 to 17 minutes, depending on download speed, you install Ollama in PowerShell (no administrator rights), pull the tag llama3.2:3b, check a French prompt, and tick the checkpoint table. The PC you already have is enough. The next section is for people who want to buy something after that proof.

Later: hardware and the selector (outside the timer)

You can do this whole path on the PC you already own. If the 3B answers and you want to move up, measure a week of real use (wait time, disk, how many people use it) before you buy.

How the LocAIhome hardware selector works. You enter a rough budget, the models you have in mind, and the household’s price per kWh. The tool suggests hardware profiles that line up with VRAM, SSD storage, and the mini PC comparison. It does not replace reading the hardware guide. Nothing here requires a purchase. It is a decision filter after your Ollama test.

Hardware signals after a week of real use
What you notice Where to look
A short reply takes more than 1 minute VRAM and model size
The disk is nearly full after 2–3 models An SSD for LLM models
A shared machine, several people Choose hardware for home AI

Price ranges and stock as of October 2026: always check the current price on the store page. A broad sample search on Amazon: mini PCs and 32 GB RAM for an LLM (Amazon affiliate link — how affiliation works).

Common questions

Why llama3.2:3b instead of a 7B model for this guide?

The 3B fits most recent PCs with 16 GB of RAM, finishes downloading in a few minutes on fiber, and answers quickly enough on CPU to prove the install works without an hour of waiting.

Does Ollama have to stay open after the installer?

Yes. A service or a taskbar icon keeps the engine running. If ollama pull fails with “connection refused,” start Ollama again from the Start menu and retry.

Can I follow this guide on a Mac or on Linux?

Yes. The same ollama commands work in a terminal after you install from ollama.com. This page spells out Windows and PowerShell (the standard install does not ask for administrator rights). On Mac or Linux, the multi-OS guide covers the differences.

Does the chat send my data to the internet?

The conversation itself stays on your PC: inference runs locally. The internet is only used for the first download of Ollama and the model, unless you later choose to open the service on the network.

Sources

Narrow it down with the hardware selector · French version of this section