How to Run AI Models on Your PC with Ollama or LM Studio

You don't need a subscription to use an AI chatbot. Free apps such as Ollama and LM Studio let you download open-weight models, such as Alibaba's Qwen and Google's Gemma, and run them on your own PC. Your conversations stay on your machine, and once a model is downloaded, it works without an internet connection.

This guide is based on Ollama's and LM Studio's official documentation and Ollama's model library, all credited at the end. Forge PC hasn't benchmarked these apps or models, so we don't quote any speed figures. How fast a model runs depends mainly on your graphics card and how much video memory (VRAM) it has.

What you need

• Windows 10 22H2 or newer. Ollama's documentation supports Windows 10 22H2 or later, Home or Pro. LM Studio supports Windows on x64 and Arm processors. On x64 PCs, LM Studio requires a processor that supports the AVX2 instruction set.

• A graphics card with plenty of VRAM. LM Studio recommends at least 4GB of dedicated VRAM. In practice, more is much better, because the whole model needs to fit in VRAM to run quickly. Our VRAM guide explains the basics.

• 16GB of system memory or more. LM Studio recommends at least 16GB.

• Storage space. Ollama needs at least 4GB for the app, and its documentation warns that models can take up "tens to hundreds of GB".

For NVIDIA cards, Ollama needs driver 551.61 or newer on Windows. For AMD cards, Ollama's Windows ROCm support covers the Radeon RX 7000 series and Radeon PRO W7000 cards. Other AMD cards, including the RX 9000 series, can use Ollama's Vulkan support instead, which is turned on by default.

Option 1: Ollama

Ollama is a lightweight app that runs models in the background. You can chat with them in Ollama's own window or from a command prompt, and other apps can connect to it.

1. Download and install. Get the Windows installer from ollama.com. It installs in your user folder and doesn't need administrator rights.

2. Choose where models are stored (optional). Models are stored in your user folder by default. To use a bigger drive, set an environment variable called OLLAMA_MODELS to a folder on that drive. In Windows 11, search Settings for "environment variables". Do this before you download any models.

3. Download and run a model. Open the Ollama app and pick a model, or open Terminal and type a command such as ollama run qwen3.5. The first run downloads the model, then a chat starts.

4. Manage your models. ollama ls lists downloaded models, ollama ps shows which are loaded, ollama stop unloads one and ollama rm deletes one.

Ollama also offers cloud models, which run on Ollama's servers rather than your PC. If you want everything to stay on your machine, choose a local model.

Option 2: LM Studio

LM Studio is a full desktop app with a chat window, which makes it easier for beginners.

1. Download and install the latest version from lmstudio.ai.

2. Download a model from the Discover tab. You can pick one of the suggested models or search for one by name.

3. Load it. Go to the Chat tab and open the model loader. Select a model, and change the load settings if you need to. Loading a model reserves the memory it needs.

4. Chat. Once the model has loaded, you can start a conversation.

LM Studio says the app can run completely offline once you've downloaded your models.

Picking a model that fits your graphics card

Most models come in several sizes, measured in billions of parameters (written as 4b, 9b, 27b and so on). Bigger models are usually more capable, but they need more memory. Here are the download sizes listed in Ollama's library on 13 November 2026:

The "Suits" column is our guide, not a guarantee. A model also needs memory for its context, which is the conversation and documents it's working with, so leave some headroom. Ollama sets a default context length based on your VRAM. It uses 4,000 tokens on cards with less than 24GB, and 32,000 on cards with 24GB to 48GB. You can raise it in Ollama's settings, but a longer context uses more memory.

Check it's using your graphics card

If a model doesn't fit in VRAM, it's split between the graphics card and system memory, and it runs much more slowly.

In Ollama, run ollama ps while a model is loaded. Ollama's documentation explains that 100% GPU means the model is entirely on the graphics card. A split such as 48%/52% CPU/GPU means part of it is in system memory. If you see a split, try a smaller model or a shorter context length.

Which graphics card should you buy for local AI?

For most people, a 16GB card is the sensible starting point. Our best GPU for local LLMs guide recommends cards such as the RTX 5060 Ti 16GB and RX 9060 XT 16GB, with 32GB options for bigger models. If you'd rather generate images than text, see our Stable Diffusion and Flux guide.

Our verdict

LM Studio is the easier way to start, with everything in one window. Ollama is lighter and better if you want other apps to use your models. Whichever you choose, pick a model that fits in your graphics card's VRAM, and use ollama ps or LM Studio's load settings to check that it isn't spilling into system memory.