Ollama Complete Beginner’s Guide (2026): Windows, PowerShell, Models, Requirements and Alternatives

Getting started: Ollama is a model runner, not a single AI model. This tutorial explains the basic concepts, the exact Windows installation and PowerShell workflow, and how to compare local alternatives without assuming that bigger models are always better.

1. What exactly is Ollama?

Imagine you want to chat with an AI assistant, but instead of sending every question to an external provider, you would like to run an AI model on your own PC. Ollama is software that downloads, starts, manages, and serves supported models. It is a model runner, not a single artificial intelligence model.

A beginner-friendly way to distinguish the pieces:

Important: Ollama is not automatically a web search engine. A downloaded offline language model cannot reliably know yesterday’s news unless you provide current information or connect a separate search/retrieval tool.

What can you do with it?

• Chat with a local assistant and ask follow-up questions.
• Draft, explain, translate, and summarize text.
• Explain code or suggest debugging steps.
• Use image input with a compatible vision model, such as gemma3:4b.
• Produce structured JSON for scripts and integrations.
• Generate embeddings used for local document search.
• Connect desktop interfaces or your own applications through its local API.

What is Ollama not?

It is not a guarantee that every model is free for every commercial use: each model has its own license. It is not a replacement for antivirus software, a backup, a factual source, or professional judgment. Running locally does not make generated answers automatically accurate.

2. Ollama’s biggest advantages and disadvantages

Bottom line: Ollama is excellent when you want local model control, repeatable automation, and privacy-conscious workflows. If you want a completely visual interface with minimal setup, LM Studio, Jan, or GPT4All may be easier for your first day.

3. System requirements: what PC do you actually need?

Official Windows software requirements

The current Ollama Windows documentation specifies Windows 10 22H2 or newer. For supported NVIDIA acceleration, it specifies drivers 551.61 or newer; AMD acceleration depends on the supported driver/backend. The normal Windows installer uses at least 4 GB for application files, and model downloads need additional disk space. See the official Windows requirements and hardware support.

Practical RAM and GPU buying guide

These are conservative starting estimates, not official minimums for every model. Exact needs depend on quantization, context length, model architecture, operating system overhead, and whether weights fit in GPU memory.

Don’t confuse download size with required RAM. A 2.5 GB model download can require substantially more available working memory once the runtime, context cache, and operating system are considered. A model that partly spills from VRAM into RAM can work but run much more slowly.

Inspect your Windows PC using PowerShell

Open Start, search PowerShell, and run:

# Windows version and installed RAM
Get-CimInstance Win32_OperatingSystem |
  Select-Object Caption, Version,
    @{Name='TotalRAM_GB';Expression={[math]::Round($_.TotalVisibleMemorySize / 1MB, 1)}}

# GPU names and driver information
Get-CimInstance Win32_VideoController |
  Select-Object Name, DriverVersion

# Free disk space on C:
Get-PSDrive -Name C | Select-Object Name, Free, Used

# NVIDIA owners: optional GPU memory / utilization check
nvidia-smi

If nvidia-smi says the command is not recognized, that does not prove Ollama is broken: you might have AMD/Intel graphics, missing NVIDIA utilities, or another supported backend.

Step-by-Step Troubleshooting and Setup

Follow the guided Beginner sequence below. Use Advanced for API configuration and performance work only after the basic setup succeeds.

Step-by-Step Troubleshooting

Follow Beginner steps to get a safe local setup working. Advanced is for API integrations and performance adjustments. Verify each step; do not expose the local AI server publicly.

BeginnerAdvanced

1

Check operating system and available memory

Confirm Windows 10 22H2 or newer, check RAM and free disk space, and identify your GPU model. Do not assume model download size equals working RAM. Start with a small model on low-memory computers.

2

Install the trusted Windows application

Download Ollama only from ollama.com/download, run the Windows installer, and launch Ollama from Start. Avoid third-party installers and unnecessary administrator rights.

3

Confirm the command is available

Open a new PowerShell window and run ollama --version. If Windows cannot find the command, reopen your terminal and inspect Ollama's installation/PATH before changing system settings.

4

Download a lightweight first model

Run ollama pull llama3.2:1b. Allow the download to complete and make sure you have enough storage. Do not attempt several large downloads simultaneously on a limited system.

5

Chat and verify a real response

Run ollama run llama3.2:1b, ask a simple question, and exit with /bye. Verify the answer is relevant; no API key is needed for local inference.

6

Verify server and model inventory

Run ollama list to see downloaded models and use Invoke-RestMethod at http://127.0.0.1:11434/api/tags to confirm the local server responds.

7

Compare one larger model if hardware permits

Try qwen3:4b or gemma3:4b, using the same neutral test prompt. Compare speed, clarity and factual accuracy before deciding that one is better.

8

Document your safe defaults

Record the working model tag, Windows version, typical response time and whether you use local-only mode. Keep the API bound to localhost and avoid exposing port 11434 publicly.

4. Install Ollama on Windows: step by step

1. Open the official Ollama download page. Avoid unofficial “cracked” installers or model bundles.
2. Download the Windows installer and run it. A standard user install does not normally require administrator privileges.
3. Launch Ollama from the Start menu. It usually starts its local server in the background.
4. Open a new PowerShell window. Run:

ollama --version

5. If a version is printed, continue. If the command is not recognized, close/reopen PowerShell and confirm Ollama is installed and running; check your user PATH rather than downloading mysterious command-line replacements.
6. Download a lightweight starting model:

ollama pull llama3.2:1b

7. Start the interactive chat:

ollama run llama3.2:1b

8. Type the following and press Enter:

Explain what DNS does to someone who has never configured a network. Use one analogy and three examples.

9. Type /bye to leave the model chat. This does not uninstall the model.

What should appear?

The first pull displays download progress. run loads the model into available memory, possibly showing an initial delay. You should then see an answer. You don’t need to set an API key for a normal local model. Optional cloud models and related integrations have different account/network requirements.

Should you run ollama serve on Windows?

Usually no. The installed Windows app runs a background server already. Running ollama serve while it is active may report that port 11434 is already in use. That generally means a server is running, not that you must reinstall Ollama. The standalone CLI/service setup is an advanced, separate scenario.

5. Choose your first local model

Catalog sizes are approximate snapshots and can change with tags/quantization. They are not RAM requirements or model-quality rankings. See Llama 3.2, Qwen3 4B, and Gemma 3.

Beginner recommendation: start with llama3.2:1b to prove installation works, then compare qwen3:4b and gemma3:4b if your memory permits. The right model is the smallest one that handles your specific tasks reliably—not automatically the largest download.

6. Essential Ollama and PowerShell commands

Be careful with ollama rm: it deletes the model from your local model list and disk storage. It does not delete your own prompts or unrelated files, but you will need to download the model again to use it.

Check whether the local API is responding

Invoke-RestMethod -Uri 'http://127.0.0.1:11434/api/tags' -Method Get |
  Select-Object -ExpandProperty models

A list of installed models is a good sign. If the connection is refused, start the Ollama app. If you see no models, try ollama pull llama3.2:1b.

7. PowerShell example: send a prompt through the local API

PowerShell can call Ollama without Python, Node.js, or a paid service.

$body = @{
    model  = 'llama3.2:1b'
    prompt = 'Give me four safe ways to troubleshoot a slow Windows PC.'
    stream = $false
} | ConvertTo-Json -Depth 5

$result = Invoke-RestMethod `
    -Uri 'http://127.0.0.1:11434/api/generate' `
    -Method Post `
    -ContentType 'application/json; charset=utf-8' `
    -Body ([System.Text.Encoding]::UTF8.GetBytes($body))

$result.response

Why stream = $false? Ollama normally streams multiple JSON messages while generating. Turning streaming off returns one JSON response, which is much easier for a beginner to use in PowerShell. Invoke-RestMethod also avoids PowerShell’s older curl alias confusion. The official API examples are available in Ollama’s API reference.

A conversational API example

$body = @{
    model = 'qwen3:4b'
    messages = @(
        @{ role = 'system'; content = 'You explain technical issues in clear English.' },
        @{ role = 'user'; content = 'What does a 404 error mean? Give three checks.' }
    )
    stream = $false
} | ConvertTo-Json -Depth 6

$response = Invoke-RestMethod `
    -Uri 'http://127.0.0.1:11434/api/chat' `
    -Method Post `
    -ContentType 'application/json; charset=utf-8' `
    -Body ([System.Text.Encoding]::UTF8.GetBytes($body))

$response.message.content

This is the foundation for building a local desktop helper, private note summarizer, or a diagnostic assistant used only by someone on the same computer.

8. Create a personalized local assistant with a Modelfile

You don’t have to repeat “explain clearly, include safety warnings” every time. A Modelfile specifies a model and default behavior. It does not retrain the model or magically give it new factual knowledge.

Create a new folder and file:

New-Item -ItemType Directory -Path "$HOME\OllamaLab" -Force | Out-Null
Set-Location "$HOME\OllamaLab"

@'
FROM qwen3:4b
PARAMETER temperature 0.3
PARAMETER num_ctx 4096
SYSTEM """You are a careful technical troubleshooting assistant. Explain the likely causes, start with safe reversible checks, warn before any action that can delete data, and say when more evidence is needed. Use plain English."""
'@ | Set-Content -Path '.\Modelfile' -Encoding UTF8

ollama pull qwen3:4b
ollama create prblem-helper -f .\Modelfile
ollama run prblem-helper

Try:

My wireless mouse disconnects twice a day. Ask me three diagnostic questions before suggesting repairs.

Why this helps: a system instruction encourages a consistent structure. It does not guarantee correct answers; always verify commands before applying them to a production computer. See the official Modelfile reference.

9. Make Ollama faster without sacrificing reliability

Optimization order matters. Trying a faster configuration before checking the actual bottleneck can make things worse.

1. Start with a smaller model. A 3B–4B model that runs fully on your GPU often feels better than a 12B model that constantly moves data between RAM and VRAM.
2. Check ollama ps. Its PROCESSOR column shows whether the loaded model runs on CPU, GPU, or a mixture.
3. Keep context modest initially. More context may require a larger KV cache and more memory.
4. Use concise input. Ask for exactly what you need instead of sending a 100-page document every time.
5. Compare two models with the same test. Judge factual accuracy and completion time, not marketing claims.
6. Avoid unnecessary simultaneous requests on limited RAM or VRAM.
7. Update GPU drivers and Ollama carefully, especially when hardware support changes.
8. Do not aggressively increase parallelism or context length until you know there is memory headroom.

Adjust context for one chat session

Inside an interactive ollama run chat, enter:

/set parameter num_ctx 4096

Or supply options through the API:

$body = @{
  model = 'qwen3:4b'
  prompt = 'Explain how to distinguish DNS failure from Wi-Fi failure.'
  stream = $false
  options = @{ num_ctx = 4096; temperature = 0.2; num_predict = 300 }
} | ConvertTo-Json -Depth 6

(Invoke-RestMethod -Uri 'http://127.0.0.1:11434/api/generate' `
  -Method Post -ContentType 'application/json' `
  -Body ([System.Text.Encoding]::UTF8.GetBytes($body))).response

The num_predict setting limits the maximum generated tokens; it does not improve the model’s underlying intelligence.

Check GPU use

ollama ps
# NVIDIA only, if the utility is installed:
nvidia-smi

Example interpretation: 100% GPU generally indicates the loaded model is using GPU memory; 100% CPU means the workload is on system memory/CPU. Your exact results depend on hardware and model support.

10. Move downloaded models to another drive safely

Models can consume significant disk space. On Windows, the default directory is inside your user profile under .ollama\models. To store new downloads elsewhere:

New-Item -ItemType Directory -Path 'D:\OllamaModels' -Force | Out-Null
[Environment]::SetEnvironmentVariable('OLLAMA_MODELS', 'D:\OllamaModels', 'User')

Then quit Ollama from its tray icon and start it again. Open a new PowerShell window, pull a small model and verify where its files appear. If you already have large existing models, do not delete the old folder until they have been safely copied/migrated and the new location has been verified. A drive move is not performed automatically just by setting the variable.

To reverse the user environment-variable setting:

[Environment]::SetEnvironmentVariable('OLLAMA_MODELS', $null, 'User')

Restart Ollama again. The Windows documentation explains model and installation directories.

11. Privacy, offline use, and local-network security

Local does not always mean offline

For an ordinary downloaded local model, questions normally execute on your PC and require no cloud inference. Downloading a model and checking for updates do use the internet. Optional Ollama cloud models, web-connected tools, or a separate chat frontend can send requests to other services.

For a strict local-only configuration, the Ollama documentation provides a cloud-disable setting:

[Environment]::SetEnvironmentVariable('OLLAMA_NO_CLOUD', '1', 'User')

Quit/restart Ollama to apply it. This disables Ollama cloud features; it does not guarantee that unrelated third-party tools or operating-system services make no network requests.

Do not expose an unprotected API publicly

The default local server address is 127.0.0.1:11434, accessible from your machine. Do not casually switch OLLAMA_HOST to 0.0.0.0 and open TCP 11434 to the internet. The local API is not a drop-in authenticated public service. For deliberate remote access, use an authenticated gateway, restrictive firewall rules, and a properly managed deployment.

Before submitting sensitive documents, also check what your chosen model license, frontend, logging system, and integrations do with the content. For truly confidential material, use organization-approved security practices.

12. Ollama alternatives: which is better for you?

These programs are not all direct equivalents: some are graphical applications, some are low-level runtimes, and some are web interfaces built on top of Ollama.

Practical recommendations:

• Just want a visual local chatbot? Start with LM Studio or Jan.
• Want a simple API, PowerShell automation, or coding integration? Start with Ollama.
• Want a web chat interface but still use Ollama models? Add Open WebUI later.
• Want to tune quantized runtimes manually? Explore llama.cpp.
• Want to serve a model to many concurrent users? Evaluate vLLM and your deployment constraints.

No tool is inherently best for everyone. Test on your own machine with the same prompts and workloads; capabilities, licensing and costs change over time.

13. Four practical examples anyone can try

Example A: explain a Windows error safely

Prompt:

I am new to troubleshooting. Windows shows “The network path was not found.” Ask which Windows version I use, whether other websites open, and whether I am using a file share. Then give three safe checks. Do not recommend registry edits or disabling the firewall as a first step.

Why this works: a precise question asks the model to collect evidence before proposing potentially risky changes.

Example B: write an email without exposing private data

Prompt:

Write a polite 100-word email asking a service provider about an outage. Use placeholders [SERVICE], [DATE], and [ERROR]. Do not invent an outage cause or include personal account information.

Why this works: reusable placeholders are safer than sending passwords or real customer records.

Example C: compare two models with a repeatable checklist

Run the same prompt against llama3.2:3b and qwen3:4b:

Explain DNS, DHCP, and NAT to a beginner. Give one accurate analogy for each, one warning about common misconceptions, and a three-question quiz.

Score both answers for accuracy, clarity, speed, and whether they followed the instructions. Repeat the test. A model that is slightly faster but consistently invents definitions is not a better choice for technical support.

Example D: summarize your own non-sensitive text file

# Create an example text file; do not feed private documents into unreviewed workflows.
'Server restarted at 10:00. Site recovered at 10:03. No data loss confirmed.' |
  Set-Content -Path '.\example-log.txt' -Encoding UTF8

$text = Get-Content '.\example-log.txt' -Raw
$body = @{
    model = 'llama3.2:1b'
    prompt = "Summarize this incident in three factual bullet points. Do not invent details:`n$text"
    stream = $false
} | ConvertTo-Json -Depth 5

(Invoke-RestMethod -Uri 'http://127.0.0.1:11434/api/generate' `
    -Method Post -ContentType 'application/json' `
    -Body ([System.Text.Encoding]::UTF8.GetBytes($body))).response

Verify the summary against the original log. A language model can still invent an explanation for the incident even if your prompt explicitly says not to.

14. Advanced features worth exploring later

Image understanding with a vision-capable model

For example, gemma3:4b supports image input. Use this to ask the model what is visible in a screenshot, then independently verify any technical diagnosis. Vision support is model-dependent: a text-only model cannot do the same task simply because Ollama has a vision feature. See the vision API documentation.

Structured JSON output

For scripting, you can request a JSON object and validate its fields in your script. A JSON-shaped answer does not guarantee the facts inside it are correct. See structured outputs.

Embeddings for private search

An embedding model converts text into numeric vectors used for similarity search. This is useful for a personal documentation search tool. It does not automatically train the chat model on your files.

ollama pull embeddinggemma

$body = @{
  model = 'embeddinggemma'
  input = 'A printer is offline in Windows 11.'
} | ConvertTo-Json

$result = Invoke-RestMethod -Uri 'http://127.0.0.1:11434/api/embed' `
  -Method Post -ContentType 'application/json' `
  -Body ([System.Text.Encoding]::UTF8.GetBytes($body))

"Embedding length: $($result.embeddings[0].Count)"

The result is a numerical vector, not a conversational response. For useful document question answering, you still need document parsing, chunking, retrieval, and evaluation. See the embedding API documentation.

15. Common Ollama problems and how to fix them

Where are the Windows logs?

In PowerShell:

explorer "$env:LOCALAPPDATA\Ollama"

Look for server.log and application logs. Remove sensitive usernames, file paths, access tokens, and prompt contents before sharing logs publicly.

16. Installation on macOS and Linux (quick reference)

macOS: Download the macOS application from ollama.com/download, move it to Applications if prompted, launch it, and open Terminal. Then try ollama --version followed by ollama run llama3.2:1b. Supported acceleration differs by computer and macOS version; use the official download and hardware documentation for current compatibility.

Linux: The official documentation provides the installation command below. Review scripts before running them with elevated privileges, use a trusted connection, and consult your distribution’s service management guidance:

curl -fsSL https://ollama.com/install.sh | sh
ollama --version
ollama pull llama3.2:1b
ollama run llama3.2:1b

A normal Linux installation may configure a systemd Ollama service. On a system where no service is running, ollama serve starts the API in the current session. Avoid accidentally starting two servers on the same port. For Linux GPU acceleration, consult the Ollama GPU reference.

---

17. Frequently asked questions

Is Ollama free?
You can run compatible local models without a per-request cloud inference fee. Your hardware, electricity, storage, particular model licenses, and optional cloud offerings still matter. Check current terms before commercial use.

Do I need an NVIDIA GPU?
No. CPU-only use is possible, and supported AMD/other GPU backends exist. GPU acceleration makes many workloads much faster but is not a universal installation requirement.

Will it work with 8 GB RAM?
Often with small quantized models, yes, depending on free memory. Start at 1B, keep prompts short, and expect CPU-only runs to be slower.

Can Ollama run without the internet?
A downloaded local model can work offline. Initial downloads and optional cloud or web features require connectivity. You can explicitly disable Ollama cloud features if needed.

Can Ollama access my files automatically?
Not merely because it runs locally. It only receives information supplied to the model or passed through configured tools/integrations. Treat integrations with file-system access as separate security decisions.

Can it replace ChatGPT, Claude, or Gemini?
It can replace some cloud-chat tasks, especially private notes, drafts, and repeated automation, but quality and features depend on the local model and hardware. Ollama is a runtime; it is not itself a direct model-to-model comparison.

Will it save money on AI credits?
If you routinely send large volumes of non-sensitive work through a paid cloud API, a local model may lower marginal API charges. However, measure electricity, hardware investment, latency, and output quality. The cheapest result is not helpful if it requires extensive manual corrections.

What should I install first?
Install Ollama, run llama3.2:1b, confirm the local API works, and then test qwen3:4b or gemma3:4b. If you dislike using a terminal, consider LM Studio or Jan before building custom tools.

Solution Checklist

0/12 completed

Download the official Ollama installer and verify Windows version and free storage.Record RAM, GPU, driver and approximate free disk capacity.Run ollama --version from a new PowerShell window.Pull a lightweight model and complete an interactive test.Verify localhost:11434/api/tags and inspect the local model list.Check whether the chosen model runs on CPU, GPU or both.Compare the same test prompt on two models and record the result.Understand the difference between model disk size and runtime RAM/VRAM.Keep the local API off the public internet; do not expose port 11434.Decide explicitly whether optional cloud features should be disabled.Check that your model's license permits your intended use.Save the working model tag and a rollback plan before major changes.

Official downloads and reference links

• Ollama official download
• Ollama Windows installation and hardware notes
• Ollama CLI reference
• Ollama FAQ, configuration and privacy
• Ollama API reference
• Ollama model library
• LM Studio and offline documentation
• Jan and Windows documentation
• GPT4All and documentation
• llama.cpp
• vLLM documentation
• Open WebUI integration with Ollama

Editorial note: This guide was reviewed on October 14, 2026. Ollama releases, driver requirements, model tags, licenses, and third-party app features change. Use the linked official documentation when making a hardware purchase or deploying a public service.