How to Run AI Locally on a Windows PC: Ollama or LM Studio

You can run a chat AI locally on Windows without learning to code. The simplest route is to install LM Studio, choose a small model, and use its graphical chat interface. If you want PowerShell, integrations, or a local API, Ollama is often the more convenient option. Both tools run downloaded model weights on your PC, but the experience depends on available memory, the model, and its quantization.

This article covers the practical Windows setup. For the broader privacy, hardware, cost, and quality questions, read our separate guide to the tradeoffs of local AI models.

Documentation and interface labels were checked on July 31, 2026. Applications, requirements, and available models change, so confirm current details on the developers’ official pages before installing.

What “running AI locally” means

In this guide, local AI means inference: your computer loads existing language-model weights into system RAM and, when possible, GPU memory, then computes the response. It is not training a neural network or building a model from scratch. You do not need a dataset, Python, or machine-learning experience for a first local chat.

A processing path is local; an application name alone is not. Prompts and outputs can remain on the PC when you use a downloaded local model without web search, a remote MCP server, or an external API. Ollama now also offers cloud models, for example: a tag containing :cloud runs remotely rather than on your GPU.

Ollama vs LM Studio: which should you choose?

Criterion	LM Studio	Ollama
Easiest first setup	Graphical interface for finding, downloading, and chatting with models	Has an app and menu, but PowerShell remains the main practical workflow
Model selection	Shows GGUF variants, quantization, and a memory estimate	Downloads a model through a short tag from the Ollama library
Local API	Enabled in Developer; default port 1234	Available after installation; default port 11434
Best fit	Beginners who want a chat interface without commands	Users building scripts, integrations, or a local service

Install only one tool for the first test. If you install both and download the same model separately in each, you may keep duplicate multi-gigabyte files on disk.

Step 1: Check Windows, RAM, VRAM, and storage

1. Press Ctrl + Shift + Esc to open Task Manager.
2. Open Performance, select Memory, and note the total system RAM.
3. Select GPU and note the dedicated GPU memory. Do not confuse it with shared GPU memory borrowed from system RAM.
4. Open Settings → System → Storage and check free space. Each downloaded model needs its own storage in addition to the application.
5. Update the GPU driver from the hardware maker’s official source.

LM Studio’s official Windows requirements recommend at least 16GB of RAM and 4GB of dedicated VRAM; x64 processors need AVX2 support. That is an application recommendation, not a promise that every large model will fit. Ollama requires Windows 10 22H2 or newer. Its Windows installer needs at least 4GB of storage, plus the space used by models.

Hardware situation	Sensible first test	What to expect
8GB RAM	A very small 2B–4B model at 4-bit, with other apps closed	It may run, but this is below LM Studio’s recommendation; speed and context will be limited
16GB RAM	A small 3B–5B model or a moderate 7B–8B 4-bit model	A practical starting configuration, not a universal guarantee
32GB RAM	A larger model or longer context after checking the memory estimate	More room for the model and normal desktop applications
Dedicated GPU	Offload part or all of a compatible model to VRAM	Usually faster when the model, driver, and backend are supported

These are first-test guidelines, not a formula that maps a parameter count to an exact RAM figure. Context, KV cache, the runtime, Windows, and open applications all consume additional memory. A model’s download size is not its total memory requirement.

Step 2: Choose your first model

• Look for instruct or chat. A base model is intended for different workflows and may follow ordinary instructions poorly.
• Start small. A responsive 3B–5B model is a better installation test than the largest model your PC can barely load.
• Start with 4-bit GGUF. A variant such as Q4_K_M is often a practical memory-quality balance. If you have headroom, compare Q5, Q6, or a larger model later.
• Test every language you need. If Ukrainian matters, do not evaluate only English prompts.
• Read the license for the exact file or tag. “Open weights” does not automatically permit every commercial use.

The commands below use qwen3:4b as a small multilingual example verified when this guide was reviewed. It is a smoke test for the setup, not a claim that this is the best model for every task. Tags, file sizes, and available versions change.

Option 1: Run local AI with LM Studio

1. Download LM Studio from its official site, install it, and launch the app.
2. Open Discover and find a small instruct/chat model from a trusted publisher.
3. Choose a GGUF variant with 4-bit quantization or higher when the memory estimate shows enough headroom. LM Studio explains that lower precision reduces file and memory requirements but can reduce quality.
4. Download the model and wait until it finishes completely.
5. Open Chat, choose the downloaded model in the model loader, and load it into memory.
6. Send a simple prompt: “In three bullet points, explain the difference between RAM and VRAM.”

If the model does not load, choose a smaller file, a 4-bit quantization, or a shorter context. If generation is too slow, check GPU offload in the load settings and close applications using substantial RAM or VRAM.

Option 2: Run local AI with Ollama

1. Download OllamaSetup.exe from the official page and install it. The normal per-user installation does not require administrator rights.
2. Close and reopen PowerShell so it can see the new command.
3. Download a local model, then launch it:

ollama pull qwen3:4b
ollama run qwen3:4b

The first command downloads the model; the second opens an interactive chat. Do not use a tag containing :cloud when your goal is local inference. To see downloaded and currently loaded models, use commands from the Ollama CLI reference:

ollama ls
ollama ps

In ollama ps, the PROCESSOR field shows whether the model is on CPU, GPU, or split between them. If PowerShell cannot find ollama, restart the terminal; if that fails, restart the Ollama app from the Windows Start menu.

Step 3: Verify that the model really works offline

1. Finish downloading the model and receive at least one response.
2. Close the chat, disable Wi-Fi, or unplug Ethernet.
3. Open a new chat with the same local model and repeat the test prompt.
4. Do not enable web search, a remote MCP server, a cloud model, or another external integration.

LM Studio states that downloaded models, local document chat, and its local server can work without an internet connection. Model search, downloads, and runtime downloads require the network. Ollama also offers a local-only switch: create a user environment variable named OLLAMA_NO_CLOUD with the value 1, then restart Ollama. This disables cloud models and web search.

An offline test confirms the specific path you tested. It does not prove that an app never checks for updates after connectivity returns. Sensitive workflows still depend on disk encryption, Windows security, backups, and permissions granted to other applications. Use our AI privacy checklist before adding documents.

Step 4: Turn the PC into a local AI service

localhost means the current computer. Another application on that PC can send HTTP requests to a local model without contacting an external server. This is one common meaning of a local AI service.

Ollama API

After installation, the Ollama API is normally available at http://localhost:11434/api. In PowerShell, check the local model list with:

Invoke-RestMethod http://localhost:11434/api/tags

LM Studio API

In LM Studio, open Developer and enable Start server. The default address is http://localhost:1234. The documentation recommends native API v1 for new integrations and also supports OpenAI-compatible endpoints. You can start the server with lms server start.

Do not expose ports 11434 or 1234 to the public internet. The local Ollama API requires no authentication, and LM Studio may also run without a token by default. Keep the bind address on localhost. If other devices intentionally need access, design firewall rules, authentication, encryption, and authorization separately; LM Studio can require an API token.

Common problems and quick checks

Problem	What to check
The ollama command is not found	Restart PowerShell and Ollama; confirm you used the official standard installer
The model does not fit in memory	Choose a smaller model or quantization, reduce context, and close heavy apps
Generation is very slow	Check ollama ps or LM Studio GPU offload; update the driver
Storage is full	Remove unused models; Ollama stores them in C:\Users\%username%\.ollama\models unless you changed the path
The offline test fails	Confirm the model and runtime finished downloading and that you did not select a cloud model or web search

A safe first workflow

Start with a small model, invented text, and an ordinary chat. Do not immediately add documents, MCP servers, agents, or folder access. MCP servers can execute code, read files, and use the network, so LM Studio advises installing only servers you trust.

After the basic test, build a set of ten to twenty real but non-sensitive tasks and compare the answers with expected results. If you want to use your own documents, a later step may be RAG, which adds retrieved passages to a prompt without retraining the model. Our separate article explains how RAG works and where it fails.

Frequently asked questions

Can you run local AI without a GPU?

Yes. Small quantized models can run on a CPU and system RAM, usually more slowly. Test a small model on the existing PC before buying new hardware.

Is 8GB of RAM enough?

Sometimes for the smallest models, but LM Studio recommends 16GB. With 8GB, close other applications, select a small 4-bit model, and keep context modest. It is enough to experiment, not a promise of comfortable performance.

Does Ollama always run locally?

No. A downloaded local model runs on the PC, but cloud models and web search use the network. Avoid :cloud tags, perform the offline test, and set OLLAMA_NO_CLOUD=1 if you want to disable cloud features.

Which is easier, Ollama or LM Studio?

LM Studio is usually easier for a command-free chat. Ollama is convenient for APIs, PowerShell, and connecting other applications. Response quality depends primarily on the model and settings rather than the interface.

What does “local machine learning services” mean?

In the context of this guide, it means a model exposed as a service or API on your own PC, such as Ollama on localhost. The phrase can also refer to an organization’s on-premises or edge deployment, or to a contractor offering machine-learning work, so its meaning depends on the search context.

Is local AI free?

An application and some models may be available without per-request fees, but the computer, electricity, storage, and maintenance still cost money. Each model’s license separately defines permitted use.

Bottom line

For a first local chat on Windows, choose one tool and one small instruct model. LM Studio provides the simpler graphical path; Ollama provides short commands and a ready local API. Check RAM, VRAM, and storage, begin with a 4-bit variant, test every language you need, and disconnect the network for a controlled offline run. Do not expose the local API or add sensitive documents until you have verified the full workflow.