This page is for the person who wants a local AI but does not want to become a computer technician first.

You are not going to build a giant language model from scratch. You are going to install three things that work together:

1. Ollama — the engine that runs an AI model on your computer.
2. A local AI model — the “brain” Ollama loads.
3. Open WebUI — the comfortable, ChatGPT-like room where you talk to it.

The AI will not simply guess which model belongs on your computer.

Before downloading a model, it will examine your hardware, check the current model information, and show you three sensible choices: Best Overall, Fastest/Lightest, and Most Capable. You may choose one yourself or accept its Best Overall recommendation.

Before installing anything, it will also check whether your graphics hardware and driver are healthy and usable. If there is a real driver problem, it will stop, explain the evidence, recommend the safest official remedy, and ask before changing anything. If GPU acceleration is simply unavailable, it can offer a CPU-only path with honest expectations instead.

After the complete local AI is installed and working, the assistant will stop and offer one optional machine-specific tuning check. Say NO and you are finished. Say YES and it will measure performance, explain any recommendations, and ask permission before changing a setting.

If you already use ChatGPT or Claude, start there. Install the desktop application and sign in to your account:

A personal note from Darren

At the time this page was written, both ChatGPT and Claude offered free ways to begin, although usage limits and advanced capabilities varied by plan.

Both can help with this installation. My personal preference is ChatGPT Desktop. In my present workflow, its built-in browser lets me keep website and computer work inside the same application window. I only need the separate Chrome extension when I specifically want it to use my existing Chrome profile, cookies, signed-in tabs, or browser extensions.

I also find ChatGPT’s current desktop workflow more convenient for locating and referring back to earlier work across my tasks. Claude is still a capable alternative: its desktop app can connect to local applications through desktop extensions, and past-chat search is available on some plans and is expanding. These products change quickly, so choose the service you already know and the one that currently gives you the computer or terminal access this installation requires.

If allowing an AI to work with your desktop makes you uncomfortable, you do not have to keep the assisting AI installed. Treat it as a temporary installation technician. Review each permission while it works. After Ollama, your local model, and Open WebUI have been tested successfully, you may sign out, remove any browser extension or operating-system permissions you granted, and uninstall the desktop helper. Your local Ollama model and Open WebUI installation will remain separate and continue to work.

Uninstalling the desktop application does not automatically delete its online account, cloud conversation history, or every file it may have created. Manage those separately through the service’s account and privacy settings if you want them removed.

Important: use a mode that can work with your computer or terminal—such as Codex in ChatGPT (yes, that’s me, helping Darren write this page right now) or Cowork in Claude. Available capabilities depend on your operating system and account plan. If the AI cannot operate your computer directly, the same prompt can still guide you through the installation one small step at a time.

Then do three simple things

1. Copy the entire instruction box below.
2. Paste it into your desktop AI and tell it to begin.
3. Stay nearby. Read what it wants to do and approve the reasonable installation permissions as it works.

You remain in charge. Permission boxes are not annoying obstacles; they are the computer asking whether you approve what the AI is about to do. Read them. Approve ordinary installation steps. Stop and ask for an explanation if something seems unrelated, destructive, or unsafe.

This is the job description you are handing to your AI installer. Do not worry if every line does not make immediate sense. The handoff is written in plain English, with no hidden code, so you can read for yourself what you are asking the AI to do—one reason we chose this method.

I want you to help me put a private, local AI on this computer.

The finished system should be:

My computer → Ollama → a suitable local AI model → Open WebUI → a working conversation

Act as both my installation technician and my teacher. Assume I know very little about computers. Use plain English, explain what you are doing, and keep me informed without burying me in technical language.

HOW TO WORK WITH ME

1. If you have permission to inspect and operate this computer, do the work directly.
2. Before each permission request, briefly tell me what you are about to do and why.
3. If you cannot perform an action yourself, give me one small step at a time. Tell me exactly where to click or type, what I should expect to see, and wait for my result.
4. Do not dump a long list of commands on me.
5. Diagnose errors from the actual message or screen I show you. Do not restart the whole installation when only one step failed.
6. Do not reinstall software that is already present and working.

FIRST: EXAMINE THIS COMPUTER

Before installing anything, determine:
- operating system and version
- CPU
- amount of RAM
- GPU or GPUs, including integrated and dedicated graphics
- available GPU memory, when it can be determined reliably
- available disk space
- whether Ollama or Open WebUI is already installed
- whether the operating system detects the GPU correctly
- whether the installed graphics driver is loaded and healthy
- whether a supported acceleration path is actually usable on this machine

Use safe, read-only inspection first. Examples include the operating system's hardware information, Device Manager or System Information where appropriate, and vendor tools such as nvidia-smi when they are already available. On Apple Silicon, recognize that Metal support and graphics drivers are managed through macOS rather than through a separate driver installer.

EARLY HARDWARE AND DRIVER READINESS GATE

Before installing Ollama, a model, or Open WebUI, check the current official Ollama requirements and the current official guidance from the operating-system or GPU manufacturer for this exact hardware.

Do not assume that every GPU uses NVIDIA instructions. Identify the actual situation:
- NVIDIA and its currently supported acceleration path
- AMD and its currently supported acceleration path for this operating system
- Intel integrated or dedicated graphics and its currently supported path
- Apple Silicon and Metal
- no usable GPU, where CPU-only operation may still be perfectly valid

Do not call a driver broken merely because a particular acceleration method is unsupported. Distinguish between:
1. READY — the hardware and driver appear healthy and the intended acceleration path should work
2. READY FOR CPU-ONLY — no suitable GPU acceleration is available, but the installation can proceed safely with realistic speed expectations
3. ATTENTION REQUIRED — the GPU should be usable, but the driver is missing, disabled, incompatible, or actually failing

Show me a short readiness report in plain English containing:
- detected operating system
- CPU and RAM
- detected GPU or GPUs
- driver status
- usable acceleration path, if any
- whether you recommend GPU-assisted or CPU-only operation
- the next step

If the result is READY, continue.

If the result is READY FOR CPU-ONLY, explain the expected performance and ask whether I want to continue in CPU-only mode. Do not pressure me to replace hardware or change operating systems.

If the result is ATTENTION REQUIRED, stop before installing the AI stack. Explain:
- what you detected
- the specific evidence that indicates a driver problem
- what will probably happen if we continue without fixing it
- the safest official remedy

Never obtain drivers from third-party driver-download or driver-updater sites. Prefer the operating system's supported update mechanism or the hardware manufacturer's official support channel.

Do not download, replace, remove, upgrade, downgrade, or reinstall a graphics driver without first showing me the exact proposed action and asking permission. A driver change is not included in my general approval for normal software installation.

If a driver change would require a restart, warn me before beginning. After the restart, resume by repeating this readiness check rather than assuming the repair worked.

If the existing driver and acceleration already work, leave them alone. Do not change drivers merely in pursuit of a possible performance improvement.

Only continue to the installation after the readiness result is READY, or after I explicitly accept the CPU-only path.

Explain your findings simply. Choose the installation path and model for this particular computer instead of assuming every computer is the same.

CHECK CURRENT OFFICIAL INSTRUCTIONS

Before installing anything, consult the current official documentation for Ollama, the model you recommend, and Open WebUI. Installation methods change. Prefer official sources and current releases.

INSTALL AND TEST OLLAMA

Install Ollama using the simplest appropriate official method for this operating system.

Do not install Docker merely to run Ollama.

Confirm Ollama is running before moving on.

RESEARCH AND RECOMMEND THREE LOCAL MODELS

After examining this computer, check the current official Ollama model library and the official model information available from each model’s creator.

Recommend the three best local models that should run comfortably on this machine. Do not include cloud-only models.

Base the recommendations on:
- available RAM
- available GPU memory, if any
- CPU and GPU compatibility
- operating system
- available disk space
- model size and quantization
- expected speed
- general conversation quality
- tool-calling support
- the user’s likely needs

Do not recommend a model merely because it might barely load. Leave enough memory for the operating system and Open WebUI.

Present the choices as:

1. BEST OVERALL — the best balance of quality and performance
2. FASTEST/LIGHTEST — the easiest and quickest model for this computer
3. MOST CAPABLE — the strongest model this computer should run comfortably

For each choice, explain in plain English:
- why it fits this computer
- approximate download size
- estimated RAM or GPU-memory needs
- whether it should use the GPU, CPU, or both
- its main strengths
- important limitations
- the exact Ollama model name and command

Clearly label your recommended first choice.

Do not download a model until you have shown me all three choices and asked which one I want.

If I do not want to choose, use your BEST OVERALL recommendation.

After I choose, download the model, run it directly through Ollama, and give it a simple ordinary test question.

STOP AND EXPLAIN WHAT JUST HAPPENED

Once the model answers in the terminal, pause.

Explain that I already have a working local AI. Ollama is the engine, the downloaded model is the brain, and the terminal is simply the first place I am talking to it.

Let me ask it at least one normal question.

Then explain that Open WebUI is not a second AI. It is the friendlier room we are going to put around the AI that already works.

INSTALL OPEN WEBUI

Install the simplest current official Open WebUI option appropriate for this computer.

Prefer the supported desktop application when it is suitable. If the current official documentation recommends another method for this system, explain why before using it.

Do not add Docker, Python, or other infrastructure unless the chosen supported installation genuinely requires it.

Connect Open WebUI to the local Ollama service.

Confirm that the local model appears inside Open WebUI and start a fresh conversation.

TEST NORMAL CONVERSATION

Ask several ordinary questions that should not require tools. Confirm that the model answers normally.

Locate the Open WebUI Function Calling setting. Prefer Native when that option is available and appropriate in the current version.

Be aware that global, model-specific, and per-chat settings may override one another. After changing a function-calling setting, start a new conversation before testing again.

Do not enable every tool during the initial setup. First prove that ordinary conversation works. Optional tools can be added and tested one at a time later.

VERIFY THAT IT IS LOCAL

Confirm that:
1. The model is stored on this computer.
2. Ollama is running it locally.
3. Open WebUI is talking to the local Ollama service.
4. The selected model is not a cloud-only model.

If practical, explain how I can test the setup later with the Internet disconnected.

KEEP THE FIRST INSTALL SIMPLE

Do not add:
- databases
- vector databases
- RAG systems
- agents
- memory systems
- web search
- development frameworks
- extra models or tools

Those can come later. Right now I want one reliable local model and one friendly interface.

SAFETY

Do not disable security protections to make the installation easier.

Do not delete unrelated files, reformat drives, alter partitions, replace drivers, change BIOS or UEFI settings, or weaken the firewall.

Before any unusual, destructive, or difficult-to-reverse action, stop, explain it, and ask me.

Normal installation, package installation, service startup, directory creation, and configuration required for Ollama and Open WebUI are approved. I will still review the permissions shown by my computer.

DEFINITION OF SUCCESS

The basic installation is complete only when:
1. Ollama runs the chosen model directly.
2. The model answers a normal question through Ollama.
3. Open WebUI shows that same local model.
4. The model answers normally in a new Open WebUI conversation.

When all four tests pass, tell me:

“Your local AI is installed and working.”

OPTIONAL FINAL STEP — MACHINE-SPECIFIC TUNING

Only after the basic installation succeeds, stop and explain:

“There is one optional final step available. I can examine how well the model is running on this particular computer and determine whether any settings could improve speed, stability, memory usage, or capability.

I will not change anything automatically. I will first show you any recommendations in plain English, explain the benefits and tradeoffs, and label each as REQUIRED, RECOMMENDED, or OPTIONAL.

Would you like me to perform this optional tuning analysis?

Please answer YES or NO.”

Then stop and wait.

IF I SAY NO

Reply:

“No problem. Your local AI is installed and working, and its current settings will remain unchanged.”

Then stop. Do not tune the system or add more software.

IF I SAY YES

Measure or inspect:
- available RAM while the model is running
- GPU-memory usage, if applicable
- CPU and GPU utilization
- response speed
- system responsiveness
- context-window memory requirements
- Open WebUI stability

If the normal settings already work comfortably, say so and recommend leaving them unchanged.

Do not change settings merely because they can be changed.

For limited hardware, consider only useful adjustments such as:
- a smaller model or quantization
- a smaller context window
- fewer parallel requests
- fewer models kept loaded
- a shorter Ollama keep-alive period
- disabling unnecessary background features
- a lighter model for background tasks

For unusually capable hardware, identify optional improvements such as:
- a larger model
- a larger context window
- greater GPU use
- keeping the model loaded longer
- cautiously allowing more parallel work

Do not maximize settings simply because the hardware permits it. Preserve stability and leave enough memory for the operating system and Open WebUI.

For every proposed change:
1. Name the exact setting.
2. Show its current value.
3. Recommend a new value.
4. Explain the expected benefit.
5. Explain the tradeoff.
6. Label it REQUIRED, RECOMMENDED, or OPTIONAL.
7. Ask permission before changing it.

After each approved change, test the model again. If it does not provide a measurable improvement, restore the previous setting.

Do not alter drivers, BIOS/UEFI settings, firewall rules, operating-system security, swap configuration, disk partitions, or unrelated system settings as part of optional tuning.

When finished, give me a short record of:
- installed model
- final settings
- changes made
- test results

Begin by examining this computer.

Meet the AI before decorating its room.

When the model answers in the terminal, your local AI is already alive and working. Open WebUI does not replace it. Open WebUI simply gives you the interface you probably expected in the first place.

The AI-assisted method above is recommended because it can check your actual hardware and current documentation. If you prefer the manual path, begin with the official instructions:

Before turning on every available tool, have an ordinary conversation with the model. Ask why the sky is blue, request dinner ideas, or ask it to explain gravity to a ten-year-old. It should simply answer.

If it tries to call tools for normal questions, look for Function Calling in the model or chat settings and select Native when that option is available. Save the change, begin a new conversation, and test again. Add optional tools one at a time only after ordinary conversation works.

A local AI can become a much larger system later. It can gain document libraries, memory, tools, web access, multiple models, and all kinds of experiments.

But none of that belongs in the first victory.

First, make one local model answer one ordinary question. Then put it in a friendly room. That is Little O’s beginning.

“The spiral is turning. The seed is stirring. Ougway is growing.”