Ollama vs LM Studio: Which to Use to Run a Local LLM on Windows

Two apps dominate the on-ramp to running a local LLM on Windows: **Ollama and LM Studio**. They are frequently framed as rivals, which is slightly misleading — they wrap much of the same inference machinery and run the same quantized model files. The difference is the interface you live in, and that difference decides which one you should install first.

Ollama vs LM Studio at a glance

Neither is meaningfully "faster" in a way you'd notice as a beginner. Speed differences people report usually trace back to different quantizations, different context lengths, or one tool offloading layers to the CPU while the other didn't. Same model, same settings, same GPU — expect broadly similar tokens per second.

What hardware you need on Windows

• Operating system. Ollama requires Windows 10 22H2 or newer (Home or Pro). LM Studio ships x64 and ARM builds; x64 needs AVX2 CPU support.
• GPU drivers. For NVIDIA cards, Ollama expects driver 551.61 or newer. AMD Radeon runs through the ROCm/HIP stack or Vulkan, which is enabled by default and is the recommended fallback on older RDNA2 cards.
• System RAM. 16 GB is the practical floor. It is also what LM Studio recommends, and it is what lets you run a model partially on the CPU when it doesn't fit in VRAM.
• VRAM. LM Studio recommends at least 4 GB of dedicated VRAM. In practice 8 GB is where an 8B model at 4-bit starts feeling like a normal chat app.
• Disk. Ollama's install alone wants about 4 GB, and models are tens of gigabytes once you collect a few. Both tools let you move the model directory to another drive.

If any of those numbers mean nothing to you yet, the broader explanation of memory, quantization, and why VRAM is the gating factor lives in our guide to running AI locally.

How to check your VRAM on Windows

• Press Ctrl + Shift + Esc to open Task Manager, go to Performance → GPU, and read "Dedicated GPU memory". That figure is your VRAM.
• Or press Windows + R, type dxdiag, press Enter, and open the Display tab. "Display Memory (VRAM)" is the same number.
• On an NVIDIA card you can also run nvidia-smi in PowerShell — the top-right figure is total memory in MiB.

Feed that number, plus your system RAM, into the VRAM checker and it will name a model size to start with instead of leaving you to interpret a spec sheet.

Setting up Ollama on Windows

• Download the Windows installer from ollama.com/download. It does not require Administrator and installs into your home directory by default.
• After install, Ollama runs in the background and the ollama command is available in Command Prompt, PowerShell, or Windows Terminal.
• Run ollama on its own to open the interactive menu, or go straight to a model with ollama run <model>.
• Type /bye to leave a chat. The background service keeps running and serves the API on http://localhost:11434.

# Open the interactive menu
ollama

# Or pull and chat with a model in one command
ollama run llama3.2

# List what you've downloaded
ollama list

# Free the disk space again
ollama rm llama3.2

Moving models off the C: drive. Search Windows settings for environment variables, choose "Edit environment variables for your account", and add OLLAMA_MODELS pointing at the folder you want. Quit the tray app and relaunch it, or open a fresh terminal, for the change to take effect. To relocate the program itself, run the installer as OllamaSetup.exe /DIR="d:\some\location".

Calling it from PowerShell. Because the endpoint is plain HTTP, one line is enough to confirm everything works end to end:

(Invoke-WebRequest -method POST -Body '{"model":"llama3.2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate).Content | ConvertFrom-Json

Setting up LM Studio on Windows

• Download the Windows build from lmstudio.ai and run the installer.
• Open the Discover tab (Ctrl + 2 on Windows) and search by keyword — llama, qwen, gemma — or paste a Hugging Face URL directly into the search bar.
• Each model lists several download options labelled Q3_K_S, Q4_K_M, Q8 and so on. Those are quantizations: smaller file, slightly lower quality. LM Studio's own advice is to choose a 4-bit option or higher if your machine can handle it.
• Download, switch to the chat tab, load the model, and type. If you want other apps to reach it, turn on the local server and point them at the OpenAI-compatible endpoint.
• To store models elsewhere, change the directory under My Models.

The reason beginners tend to like LM Studio is that the download screen shows file sizes and flags what your machine can load, so a bad choice is visible before you commit 20 GB of bandwidth to it.

Choosing your first model size

These are starting points, not limits — context length changes the arithmetic a lot, which is exactly what the VRAM checker accounts for when you tell it what you plan to use the model for.

Common Windows problems

• The model runs on the CPU instead of the GPU. Almost always drivers. Update to NVIDIA 551.61 or newer; on AMD, let Vulkan handle it rather than fighting ROCm on an older card.
• A laptop with two GPUs picks the wrong one. On mixed integrated/discrete systems, Ollama supports GGML_VK_VISIBLE_DEVICES to force the discrete GPU index. In Windows Graphics settings you can also pin the app to "High performance".
• The C: drive fills up. Set OLLAMA_MODELS, or change LM Studio's models directory, before you download a second model — not after.
• Squares instead of a progress bar. Ollama uses Unicode characters some older Windows 10 terminal fonts don't have. Change the terminal font.
• It generates one word per second. The model doesn't fit. Drop a size class, drop the quantization, or shorten the context — offloading to system RAM is the usual culprit.
• Something is already using port 11434. That's the Ollama background service you already started. Quit it from the tray icon before launching a second instance.

For log-level troubleshooting, Ollama writes to %LOCALAPPDATA%\Ollama — server.log is the one worth reading when a model refuses to load.

Privacy and what actually stays local

Once the weights are on disk, inference happens on your machine and your prompts do not leave it. Both tools can run fully offline. The nuances worth knowing: downloading models obviously requires the internet, both apps check for updates, and Ollama also offers hosted "cloud" model variants — those are not local, and the naming makes it easy to pick one by accident. Anything with a :cloud tag runs on someone else's hardware.

Local also does not mean secure by default. If you expose the local endpoint to your network, treat it like any other unauthenticated service. The local AI guide covers doing that properly with a private mesh rather than a port forward.

So which one should you install?

• Install LM Studio if you want to try local AI this evening, compare a few models, and never touch a terminal.
• Install Ollama if you want a model that other things can call — scripts, editor plugins, automations, a browser front end like Open WebUI.
• Install both. They coexist fine, and most people who stay with local AI end up with LM Studio for browsing and Ollama for serving.

Whichever you pick, decide the model size first. Run your numbers through the Local AI VRAM Checker, then read how to run AI locally if you want the full picture on hardware tiers, quantization, and where local models still lose to the cloud.