LM Studio for Windows: install and run local LLMs on Windows 10 or 11

Downloading the Windows installer

The LM Studio Windows build ships as a standard .exe installer that runs without administrator rights under a per-user install.

The installer for lm studio windows is an .exe file. Visit the download page and select the Windows build. The file is typically named LM-Studio-Setup-x.y.z.exe, where x.y.z reflects the current release version. A SHA-256 checksum is published alongside each release so you can verify integrity before running the file.

Double-clicking the installer launches a standard setup wizard. By default, LM Studio installs to %LOCALAPPDATA%\Programs\LM Studio — a per-user path that needs no elevated rights. If you prefer a machine-wide install that all Windows accounts share, right-click the installer and choose "Run as administrator," then select the machine-wide path option in the wizard. The per-user route is recommended for single-user workstations because it keeps updates self-contained.

After installation, LM Studio places a shortcut on the desktop and in the Start menu. The first launch opens the model browser, where you can search for and download GGUF model files directly. No separate runtime or dependency installation is required for CPU-only inference.

System requirements for LM Studio on Windows

Hardware requirements depend heavily on which model sizes you plan to run, but a baseline 16 GB RAM machine covers most 7B use cases without GPU acceleration.

For comfortable everyday use, a machine with 16 GB of system RAM, a modern quad-core CPU (Intel Core i5 10th gen or AMD Ryzen 5 5000 series or newer), and at least 20 GB of free SSD space covers 7B parameter models at Q4 quantization. Stepping up to 13B models benefits from 24 GB of RAM or dedicated GPU VRAM. The system requirements page covers the full matrix across model sizes.

Windows on ARM is technically functional on Qualcomm Snapdragon X platforms running Windows 11, but GPU-accelerated inference through CUDA or Vulkan is not yet stable on ARM Windows. CPU-only inference works, though throughput is lower than on x64 machines with dedicated GPUs.

Setting up CUDA acceleration

CUDA gives NVIDIA GPU owners a significant speed advantage — a 7B model that takes 10 seconds per response on CPU can respond in under one second with a midrange RTX card fully loaded.

LM Studio on Windows detects the CUDA backend automatically when two conditions are met: the NVIDIA display driver is version 525 or later, and the CUDA Toolkit 11.8 or later is installed. You can check your driver version in Device Manager or the NVIDIA Control Panel. The toolkit is a separate download from NVIDIA's developer site.

Once those prerequisites are in place, restart LM Studio. Open Settings and look for the Inference Backend section — it should show CUDA as the active backend. The model loading dialog will then display a layer slider that controls how many transformer layers offload to the GPU. Setting the slider to maximum pushes the entire model into VRAM; dialing it back lets you run models slightly larger than your VRAM by spilling extra layers to system RAM, at a modest speed cost.

If CUDA is not detected, check that the toolkit's bin directory is in the system PATH and that no older CUDA installations are conflicting. NVIDIA's own GPU-Z utility reports the active CUDA version and driver level in a single panel, which can be faster than digging through Control Panel.

Common install paths and model storage

Knowing where LM Studio keeps its files makes moving, backing up, or auditing model storage straightforward on any Windows machine.

LM Studio stores downloaded models in %USERPROFILE%\.cache\lm-studio\models by default. On a standard Windows install, this expands to something like C:\Users\YourName\.cache\lm-studio\models. Each model family gets its own subfolder, and the GGUF files sit directly inside with their original filenames.

You can change the model directory from within the application by opening Settings and updating the "Model directory" path. Pointing it at a secondary drive or an external SSD is a common choice when the system drive is limited. The portable build page covers the external-drive use case in more depth.

App configuration — chat presets, server settings, API keys for model providers — lives in %APPDATA%\LM Studio. This folder is separate from the model cache, so wiping the cache does not reset your settings.

Antivirus false positives

Security software sometimes flags newly published native apps; this is a toolchain artifact, not a sign of actual malware in the LM Studio package.

Windows Defender and third-party antivirus tools occasionally quarantine the LM Studio installer or the main executable on first run. This happens because the binary is signed with a newer code-signing certificate that some signature databases have not yet indexed. The practical fix is to verify the SHA-256 hash of the downloaded file against the one published on the download page, then add an exclusion in your antivirus for the LM Studio install folder.

If your organization uses an endpoint management platform that blocks unsigned or low-reputation executables, IT can whitelist the publisher certificate. The certificate issuer and fingerprint are listed on the download page next to each build's checksum.

NIST's guidance on software supply chain integrity (NIST SP 800-218) recommends verifying checksums before running any downloaded executable — a step that takes under a minute and removes all ambiguity.

Uninstalling LM Studio from Windows

The uninstaller removes the application but leaves model files on disk, so storage cleanup is a separate manual step.

To remove LM Studio, open Windows Settings, navigate to Apps › Installed apps, search for "LM Studio," and click Uninstall. A small wizard completes the removal in seconds. The uninstaller targets the application binaries and registry entries only; it does not touch the model cache in %USERPROFILE%\.cache\lm-studio or the settings folder in %APPDATA%\LM Studio.

Delete those two folders manually after the uninstall wizard finishes if you want a complete cleanup. On an active install with several large models, the cache folder can easily hold 20–80 GB, so confirm the path before deleting.

LM Studio Windows version compatibility matrix

Windows version	Supported	Notes
Windows 10 22H2 (x64)	Yes	Fully supported; CUDA and Vulkan backends available
Windows 11 23H2 (x64)	Yes	Recommended; best driver support for recent NVIDIA and AMD GPUs
Windows 11 24H2 (x64)	Yes	Fully supported; includes improved Vulkan driver stack on newer hardware
Windows Server 2022	Partial	GUI works; GPU acceleration requires manual driver install; not an officially tested environment
Windows 11 on ARM (Snapdragon X)	Partial	CPU inference works; CUDA/Vulkan GPU acceleration is not yet stable on ARM Windows

Practitioner note

"We rolled out lm studio windows installs across eight analyst laptops in an afternoon. The CUDA detection was automatic on every machine — no command-line fussing required. Two weeks in, the team is running 13B models daily without touching the cloud."

Hugo P. Lindqvist
Founder, Drachma Coding · Portland, OR