LM Studio GPU Not Detected: 5 Fixes That Actually Work (2026)
You loaded a model, hit send, and the response is crawling out one word every few seconds. You check Task Manager — your GPU is sitting at 2%. Your CPU, meanwhile, is pinned at 100%. Something is very wrong, and it's not your hardware.
This is almost always the same problem wearing three different disguises, and none of them require a reinstall. Here's the actual fix order, tested across NVIDIA, AMD, and Apple Silicon.
Quick Answer
If LM Studio isn't using your GPU, it's almost always one of these three things, in this order of likelihood:
- The GPU Offload slider is set to 0
- The wrong inference engine (runtime) is installed for your hardware
- Your GPU driver is out of date
Fix them in that order and you'll be back to full GPU speed in about ten minutes. Let's go through each one properly.
Fix 1: Check the GPU Offload Slider First
This single setting causes more "GPU not detected" reports than every driver issue combined, because it's easy to miss and LM Studio doesn't always default it to max.
- Load a model in the chat window.
- Click the gear icon next to the loaded model name (or open the right-hand settings panel).
- Find GPU Offload and check the number of layers assigned to GPU.
- Drag it to Max, or manually set it to match your model's total layer count.
If this was set to 0, you'll immediately see your GPU spike in Task Manager (or nvidia-smi/rocm-smi) the next time you send a message.
Why this happens: LM Studio sometimes resets GPU Offload to a conservative default after an update, or when a new model is loaded for the first time. It's not a bug so much as a setting that silently reverts more often than it should.
Fix 2: Confirm the Right Runtime Is Installed
LM Studio doesn't run inference itself — it's a frontend over llama.cpp (and MLX on Apple Silicon). Each hardware vendor needs its own compiled runtime, and if the wrong one is installed, LM Studio quietly falls back to CPU with no error message.
Go to Settings → Runtime and confirm you have the correct engine installed for your card:
| Your Hardware | Runtime You Need |
|---|---|
| NVIDIA GPU | CUDA llama.cpp |
| AMD GPU | ROCm llama.cpp (or Vulkan as a fallback) |
| Intel GPU | Vulkan llama.cpp |
| Apple Silicon (M1–M4) | MLX |
If the wrong one is selected — or if it downloaded but failed silently — switch to the correct engine and restart the app. This single mismatch is responsible for a huge share of "0 GPUs detected" reports, especially on fresh installs where the auto-detection logic guessed wrong.
Fix 3: Update Your GPU Driver (Not Just Windows Update)
If Fixes 1 and 2 didn't work, your driver is the next suspect — and specifically, a driver installed through Windows Update is often not enough.
- NVIDIA: Download the latest "Game Ready" or "Studio" driver directly from NVIDIA's website, not through Windows Update. Restart after installing.
- AMD: Make sure you have the correct ROCm version installed for your specific distribution and GPU generation — ROCm support is more hardware-specific than CUDA, so double-check your card is actually on the supported list before assuming the driver itself is the problem.
- Linux (any GPU): Confirm
nvidia-smiorrocm-smiactually detects the card at the OS level before blaming LM Studio. If your OS can't see the GPU, no app will.
Fix 4: Clear a Corrupted Local Config Cache
Less common, but worth knowing: a corrupted local settings cache can leave the runtime selection or GPU detection logic in a bad state even after you've fixed the settings above. If the GPU still won't engage:
- Quit LM Studio completely.
- Delete the local config folder — Windows:
%APPDATA%\LM Studio, or macOS/Linux:~/.config/LM Studio(or~/Library/Application Support/LM Studioon macOS). - Relaunch LM Studio and reconfigure your runtime and GPU Offload settings from scratch.
This resets LM Studio's memory of which engine it picked, forcing it to re-detect your hardware cleanly.
Fix 5: Verify With the lms CLI Instead of Guessing
LM Studio ships a command-line tool that tells you exactly what's happening under the hood, instead of you guessing from the UI.
npx lmstudio install-cli
Run this once to add lms to your system PATH (open a new terminal window afterward). Then:
lms ps
This reports the currently loaded model and — critically — how many layers are actually running on GPU versus CPU. If it shows 0 GPU layers even after you've set GPU Offload to Max, you know the runtime (Fix 2) is still the real culprit, not the slider.
lms status
Confirms whether the app and local server are running at all — useful if you're also using LM Studio's API server and things seem to have silently stopped.
Cheat Sheet
| Symptom | Likely Cause | Fix |
|---|---|---|
| GPU Offload slider shows 0 | Setting reverted or never set | Drag slider to Max in model settings |
| Correct slider, still 0% GPU usage | Wrong engine installed | Settings → Runtime → pick CUDA/ROCm/Vulkan/MLX |
| Runtime looks correct, still CPU-only | Outdated driver | Install driver directly from GPU vendor's site, not Windows Update |
| Fixed everything above, still broken | Corrupted local config | Delete LM Studio's config folder and reconfigure |
| Not sure what's actually happening | — | Run lms ps to see real GPU layer allocation |
Final Advice
Don't jump straight to reinstalling LM Studio or reflashing drivers the moment your GPU doesn't light up — in the overwhelming majority of cases it's the GPU Offload slider sitting at 0 or the wrong runtime silently installed, both of which take under a minute to fix once you know where to look. Work through the five fixes in order above rather than randomly, since each one rules out a specific layer of the stack (setting → engine → driver → cache → verification).
Once your GPU is actually being used, it's worth revisiting which model you're running in the first place — our breakdown of LM Studio Bionic and which models are worth running behind it covers model picks by VRAM tier, and if you're deciding between LM Studio and Ollama entirely, this comparison walks through which one fits your workflow better.