LM Studio
A desktop tool for downloading, loading and running large language models locally, with a chat interface, model library and compatible local API service.
What LM Studio is for
LM Studio provides a desktop workflow for finding, downloading, loading and testing local language models. The application is the runtime and interface; model weights, quantization, licenses and hardware requirements remain separate decisions.
Match model size to the device
Available memory and graphics memory matter more than free disk space alone. A model may fit on disk but fail during loading, swap heavily or become unusably slow when the context window and concurrent requests are added. Start with a smaller quantized model and a short test prompt.
Keep local services local by default
The chat interface and compatible API can expose different data paths. Verify the endpoint, model identifier, listening address and logs before connecting another client. A local model is local only when the process, plugins, attachments and integrations remain within the intended boundary.
Maintenance note
This page covers LM Studio setup, model memory planning, local chat and API exposure boundaries. Content review date: 2026-08-23.
Save to your cloud drive
Open the cloud drive to get the file directly, or save it for convenient access on another device.
Quark Cloud Drive
RecommendedSave LM Studio to this cloud drive
Baidu Netdisk
Save LM Studio to this cloud drive
LM Studio installation, model selection and local API setup
Install the matching package, test a small model locally and inspect memory and endpoint settings before connecting other applications.
Before you start
- Confirm the operating system, processor architecture, graphics support and available memory before choosing a build.
- Reserve space for model files, caches and logs; a model can occupy several gigabytes or more.
- Use non-sensitive test prompts and a disposable model for the first run.
Installation steps
- 01
Select the platform build
Choose the package for Windows, macOS or Linux and confirm the architecture. Do not force an incompatible build when the runtime or driver is unsupported.
- 02
Set the model storage location
Launch the application, check the model directory and verify that the disk has enough space for downloads, cache files and future versions.
- 03
Check hardware detection
Review available CPU, GPU and runtime options, then run a low-load test before downloading a large model.
Quick start
- 01
Choose a small model
Compare parameter size, quantization, context length, license and intended use. Start with a model that leaves memory headroom.
- 02
Load and test a conversation
Load the model, send a short non-sensitive prompt and record response speed, memory use and the displayed model identifier.
- 03
Start the local API only when needed
Enable the compatible API after chat works, keep the listener local and test one client request before changing network exposure.
Usage tips
- A model that loads successfully may still exceed practical memory during long context or concurrent requests.
- Read each model license and usage restriction separately from the application license.
- If the API binds to a LAN address, add authentication, firewall rules and a clear access boundary before sharing it.
Troubleshooting and uninstall
Why does loading stop or fail halfway through?
Close other memory-heavy programs, shorten the context and choose a smaller or more compressed model. Confirm that the model file completed downloading.
Why is generation very slow?
Check whether the intended GPU runtime is active, whether the model exceeds graphics memory and whether the context is too long. A smaller model often gives a more stable baseline.
- Decide which models to retainRecord the model names, licenses and configuration needed for future work, then remove unused copies only after confirming the active models.
- Remove the application and cacheUninstall through the operating system and clear model or cache directories separately, preserving projects and logs that are still required.
Frequently asked questions
Does LM Studio include model weights after installation?
The application and model files are separate. Download a compatible model and check its size, quantization and license before loading it.
How large a model should a first-time user choose?
Start with a smaller quantized model that leaves memory headroom, then increase context or model size only after the local workflow is stable.
Is a local API automatically protected from network access?
No. Verify its listening address, firewall, proxy and authentication settings. Keep it bound to the local machine unless a controlled network use case is required.