You've spent the better part of an hour downloading a model, and now the load bar sits at zero. I've had that evening more than once with LM Studio. The app won't open, or opens to a white window, or the API says nothing back. The symptoms look alike, but the causes rarely are.
What helped me most was deciding which stage is stuck before touching anything. Won't start, won't load, and API-not-responding are three different problems, and each has its own place to look.
Which stage is stuck?
I used to reinstall the moment something broke. It didn't go well: I'd wipe settings and models along with the problem, and end up with a different environment but no explanation. These days I sort it like this.
| Symptom | Suspect | Start with |
|---|---|---|
| Nothing happens when I click the icon | OS security, missing runtime | Step 2 (Mac) / Step 3 (Windows) |
| Opens to a white or transparent window | GPU driver, rendering | Step 1, then a driver update |
| Model won't load | Disk space, memory, a broken download | Step 4 |
| API doesn't respond | Server not running, port conflict | Step 5 |
Step 1: Update to the latest version
LM Studio ships updates often, and load-related bugs are frequently fixed in the next release. Check the official site first. If the in-app update prompt never appears, download the installer again and install over the existing copy.
I don't recommend deleting the settings folder as a first move. You lose history and presets, and before you've confirmed the settings are the cause, you also throw away your clues.
Step 2: macOS says the app "is damaged and can't be opened"
A quarantine attribute on the downloaded app is often the cause. Clear it, then open the app again.
xattr -cr "/Applications/LM Studio.app"If the app lives somewhere other than /Applications, adjust the path. If you'd rather avoid the command, "Open Anyway" under Privacy & Security in System Settings does the same job.
Step 3: Windows, nothing appears at launch
If the app exits right away, the Visual C++ Redistributable may be missing or old. List what's installed:
Get-ItemProperty HKLM:\Software\Microsoft\Windows\CurrentVersion\Uninstall\* |
Where-Object { $_.DisplayName -like "*Visual C++*" } |
Select-Object DisplayName, DisplayVersionIf nothing shows up, or the versions are old, install the latest redistributable from Microsoft. If you get a white window instead, update your graphics driver and try again.
Step 4: The model won't load
Nearly every load failure comes down to disk space, memory, or a download that was cut off. I check them in that order.
Disk space. Models run from a few GB to tens of GB. Make sure the drive holding them has comfortable headroom.
A truncated file. The model folder is whatever you set in the app's My Models tab. If you've lost track of it, ask the CLI:
lms lsIf a model's listed size is absurdly small, the download was cut off. Delete it and fetch it again.
Memory. For a 4-bit quantized model, I estimate roughly half the parameter count in GB, plus headroom for the context window. That's a rough guide only; architecture and context length shift it a lot. If memory is short, a smaller model or a more aggressive quantization often loads.
Step 5: The API doesn't respond
LM Studio's local server exposes an OpenAI-compatible API. First, check whether it's running:
lms status
curl http://localhost:1234/v1/modelsIf the lms command isn't found, launch LM Studio once, then run ~/.lmstudio/bin/lms bootstrap (on Windows, lms.exe bootstrap in the same folder, per the official instructions) and open a new terminal.
If the server is stopped, start it from the Developer (Local Server) screen in the app, or run:
lms server startIf it's running but still silent, see whether another process holds the port:
lsof -i :1234If something else is using it, change the port in LM Studio's server settings and point your curl at the same number. On Windows, also check that the firewall isn't blocking inbound connections.
Step 6: Read the logs
If you still can't see the cause, read the logs. The screen may only say "failed," while the log usually says why.
lms log stream --source serverThis keeps streaming server activity. Load the model or hit the API again while it runs, and you can follow just that moment. To see the exact strings sent to the model, use --source model instead. Server log streaming needs version 0.3.26 or later, so on an older build, go back to Step 1 first.
Paste the error text into a search engine or the issues of LM Studio's official repository; you may find someone who hit the same thing.
Confirm that it works
Last, call the model list and a chat completion once each with curl:
curl http://localhost:1234/v1/models
curl http://localhost:1234/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "a-model-name-from-the-list", "messages": [{"role": "user", "content": "Hello"}], "max_tokens": 50}'If both return JSON, everything from startup to the API is connected.
Keeping this from happening again
When I run local models day to day as an indie developer, most stalls came down to disk space and falling behind on updates. I keep two habits:
- Leave headroom in the model folder. Check free space before pulling a new model.
- Run
lms statusbefore and after a big update. If it breaks afterward, you immediately know which side to blame.
If security software is interfering with loading, you can exclude LM Studio's install folder and model folder. I'd still work through the steps above first, so you don't pile on settings you didn't need.
I've also written about day-to-day use in Working LM Studio into an Antigravity daily workflow.
Before reinstalling, narrow it to one stage. That's a rule I try to keep hardest on the nights I'm most rushed. Start with two lines, lms status and curl http://localhost:1234/v1/models.