If LM Studio isn’t detecting your GPU, the cause is almost always one of three things: the GPU Offload slider is set to 0, the wrong inference engine is installed, or your graphics driver is out of date. Fix those in that order and your card will start pulling its weight. Most people are back to fast, GPU-powered inference in about ten minutes.
The confusing part is that “not detected” can mean two different things. Sometimes the app really sees zero GPUs. Other times it sees the card fine but still runs everything on your CPU because a setting told it to. This guide covers both, starting with the fix that solves the largest share of cases.
Start here: the GPU Offload slider
Before touching drivers, check the one control that actually decides whether your GPU gets used. Load any model, click the gear icon next to its name, and find the GPU Offload slider. If it sits at 0, every layer runs on your CPU no matter how good your card is. Drag it up, or set it to Max, then reload the model.
This slider controls how many of the model’s layers move onto the GPU. At 0, none do. At Max, all of them do, as long as they fit in your video memory. This single setting is behind a large share of “my GPU isn’t being used” reports, so it is worth ruling out first.
Quick test: load a model, open the gear menu, set GPU Offload to Max, and reload. If your GPU still isn’t touched, the problem is deeper, and the sections below will find it.
Why LM Studio isn’t seeing your GPU
When the app genuinely reports zero GPUs, three root causes explain nearly all of it. Work through them top to bottom.
The wrong engine is installed. LM Studio ships as two separate pieces: the app you see, and the runtime that actually runs models. The runtime is the llama.cpp or MLX engine, and it downloads on its own schedule. A fresh install can quietly default to a CPU-only build, which means the GPU Offload slider has nothing to act on.
The driver is too old (or, rarely, too new). LM Studio’s bundled engine expects a reasonably current CUDA or Vulkan runtime underneath it. A driver more than about a year old is a common reason GPU initialization fails outright.
The hardware or platform isn’t fully supported. Cards with very little video memory, brand-new architectures the current engine wasn’t built for, and integrated graphics that get skipped when a dedicated card is present all show up as detection problems. These are the edge cases, and each has its own section below.
There’s a reason this trips up so many people. LM Studio runs models through llama.cpp under the hood, and that engine needs a GPU backend compiled specifically for your hardware. The detection chain on Windows is genuinely fragile: an NVIDIA card needs a driver version that lines up with the engine build, and a driver that’s too old or occasionally too new can break it. AMD needs Vulkan present. Intel Arc and integrated chips are the ones most often skipped even with current drivers. If you’ve spent an hour swapping driver versions with no luck, you’re not doing anything wrong. The moving parts just have to line up, and this guide lines them up in the order that matters.
Install the right engine for your hardware
This is the fix people miss most, so it’s worth doing carefully. Open Settings, then Runtime (in older builds it lives under Developer, then Runtimes; the shortcut is Ctrl+Shift+R on Windows and Linux). Install and select the engine that matches your card, then restart the app and reload your model.
| Your hardware | Engine to install | Notes |
|---|---|---|
| NVIDIA (Windows/Linux) | CUDA 12 llama.cpp | The most mature path. Auto-upgrades on a current driver. |
| AMD (Windows) | Vulkan llama.cpp | ROCm is Linux-only, so Vulkan is the Windows route. |
| AMD (Linux, RX 7000 / some 6000) | ROCm llama.cpp | Faster than Vulkan on supported cards. |
| Intel Arc / integrated | Vulkan llama.cpp | Vulkan covers most Arc and integrated chips. |
| Apple Silicon (M-series) | MLX | Default on Mac. Uses unified memory. Leave it. |
| No GPU | CPU llama.cpp | Works, but slow on anything above 7B. |
After installing, look for a green “Latest version” label next to the engine. If there’s an update button, use it. The underlying llama.cpp updates often, and newer builds can be meaningfully faster. LM Studio bundles its own CUDA runtime, so on NVIDIA you do not need to install the CUDA Toolkit separately.
The official system requirements page is worth a glance here too. On Windows and Linux your CPU must support the AVX2 instruction set, and at least 4GB of dedicated video memory is recommended. Almost every Intel chip from 2013 onward and every AMD Ryzen chip clears the AVX2 bar, but old or virtualized machines sometimes don’t.
One more thing to know about the runtime: it updates separately from the app, which is easy to forget. The app can happily update itself while your GPU engine stays stuck on an old or CPU-only build. So when a new LM Studio version lands, it’s worth reopening the Runtime tab and checking that your engine is also current. The speed gains between engine versions are often larger than the app changes themselves, and a recent build can be the difference between a card that loads a model slowly and one that loads it in a couple of seconds.
Update your graphics driver
If the right engine is installed and the slider is up but your GPU still isn’t found, the driver is the next suspect. A clean reinstall clears out corrupted or stale components that block detection.
- NVIDIA: Download the latest Game Ready or Studio driver from NVIDIA, choose Custom Installation, and tick “Perform a clean installation.” A driver from the 535 series or newer is a safe baseline.
- AMD: Install the newest Adrenalin driver. On Windows this also bundles the Vulkan runtime the engine needs.
- Then reboot. Restart the machine, not just the app. Reopen LM Studio and reload your model.
Don’t lean on Windows Device Manager for this. It often installs an older driver than the one on the manufacturer’s site. If you’re on AMD and the app shows no GPU at all, the Vulkan runtime may be missing; updating the Adrenalin driver usually restores it, or you can install it on its own from LunarG.
Fix it from the command line
If the graphical slider seems stuck or you’re automating things, the lms CLI gives you direct control. It’s the real binary name, not “lm-studio.” Start by checking what’s loaded:
lms status and lms ps
If lms ps shows something like 0/33 layers on GPU, you’re running on CPU. Load a model with full offload instead:
lms load qwen2.5-7b-instruct –gpu max
The –gpu flag also takes auto (let the app decide), a fraction like 0.5 (roughly half the layers), or off (force CPU, handy on locked-down work machines). There is no –diagnose or –no-gpu flag, so don’t reach for those.
If nothing has worked and the runtime selection seems stuck, a corrupted settings cache can be the culprit. Quit the app fully, then remove its config folder. On Windows that’s %APPDATA%\LM Studio; on macOS it’s the LM Studio folder under Library, Application Support; on Linux it lives under your home config directory. Reopen the app, reinstall your engine, and set up offload again. This resets the app to a clean state without deleting your downloaded models, which live in a separate folder.
The CLI is also the fastest way to sanity-check a headless or scripted setup. If you’re running LM Studio’s local server for a coding tool or automation and can’t watch a slider, lms ps tells you at a glance how many layers landed on the GPU. Seeing the full layer count there, rather than 0, confirms the card is doing the work before you point anything at the endpoint.
Special cases worth knowing
A few situations don’t fit the standard checklist. If you’re in one of these, start here.
Integrated GPU ignored when a dedicated card exists
On a laptop or desktop with both an integrated chip and a dedicated card, the Vulkan engine tends to lock onto the dedicated GPU and skip the integrated one entirely. This is a known quirk rather than a setting you flipped. If you specifically want the integrated GPU used, your cleanest option is to restrict the app’s access to the dedicated card at the system level, though for most people the dedicated card is the one you want anyway. The situation to watch for is a machine where the dedicated card has very little memory and the integrated chip shares a large pool of system RAM. In that narrow case the integrated path can actually load a bigger model, and the workaround above is worth the effort.
Brand-new NVIDIA cards (RTX 50-series)
A very new card can load fine on the driver side yet throw a “compute capability” error, because the active runtime was built before your architecture existed. For Blackwell cards like the RTX 5090, update the app itself to 0.3.15 or later, update your NVIDIA driver, then open the Runtimes tab and make sure the CUDA 12 llama.cpp engine is selected and current. With a new enough driver, LM Studio switches to that runtime on its own.
Two GPUs, only one used
On a dual-card rig you may see one GPU maxed out and the other idling at 0 percent. LM Studio has real multi-GPU controls: open them, confirm both cards are toggled on, and set the priority order you want. For a model too big for a single card’s memory, tensor parallelism (added in recent 0.4.x builds) splits it across both cards instead of spilling onto the CPU. If a card that should be active isn’t even listed, update the app, since older builds had a selection bug.
Linux and virtual machines
On Linux, make sure your GPU drivers and the matching compute stack are installed before you expect detection to work, and note that Ubuntu releases much newer than 22.04 are not heavily tested. In a virtual machine, GPU passthrough is often the missing piece, and a tiny VM will fail for a simpler reason: models need real memory. A box with a couple of gigabytes of RAM can’t load much of anything, GPU or not.
Check the GPU is actually working
Seeing your card listed isn’t proof it’s doing the work. Confirm real usage during inference, not just detection.
- Windows: Open Task Manager, go to Performance, and watch the GPU while a model generates text. A card that’s working shows a clear jump.
- Linux: Run nvidia-smi -l 1 in a terminal to see live utilization tick up.
- Inside LM Studio: Watch your tokens-per-second. Moving from CPU to GPU on a 7B model is a large, obvious speed jump, not a subtle one.
If the GPU is listed and offload is set high but tokens stay slow, you may be running a model too large for your video memory, which forces layers back onto the CPU. Try a smaller quantized version, such as a Q4_K_M build, or lower the offload a notch so what does fit runs cleanly on the card.
A rough memory rule helps you pick a model that fits before you even load it. A 7B model at Q4 needs around 5GB, a 13B needs about 9GB, and a 30B jumps to roughly 18GB. Compare that to your card’s video memory and leave a little headroom for the context window, which grows with how much text you feed the model. If a model’s total sits just above your memory, that’s exactly when you see a card that’s detected but crawling: a handful of layers spill to the CPU and drag the whole thing down. Dropping one quantization level down, or trimming the context length, often puts everything back on the GPU and restores full speed.
Closing other programs that lean on the GPU helps too. A browser with hardware acceleration, a game, or a video editor can quietly hold onto a chunk of video memory, leaving less for your model than the raw spec suggests. If a model that should fit refuses to load fully onto the card, shutting those down and reloading is a quick thing to try.
Frequently asked questions
Why does LM Studio say “0 GPU detected” when I have a graphics card?
Usually the matching engine was never installed, or your driver is too old. Open Settings, then Runtime, install the CUDA (NVIDIA) or Vulkan (AMD/Intel) engine, update your driver, restart, and reload the model.
Does LM Studio support AMD and Intel GPUs?
Yes. It’s no longer NVIDIA and Apple only. AMD and Intel cards run through the Vulkan engine on Windows, and some AMD cards on Linux can use the faster ROCm engine instead.
Where is the GPU setting in LM Studio?
Load a model, click the gear icon next to its name, and find the GPU Offload slider. Set it above 0, or to Max, to move model layers onto your GPU. At 0 it runs entirely on the CPU.
Do I need to install CUDA separately for LM Studio?
No. LM Studio bundles its own CUDA runtime, so you don’t need the CUDA Toolkit. You just need a reasonably current NVIDIA driver, ideally the 535 series or newer, installed with a clean installation.
My GPU is detected but LM Studio is still slow. Why?
The model is likely too big for your video memory, so layers spill back to the CPU. Try a smaller quantized version like Q4_K_M, or lower the GPU Offload so the layers that fit run cleanly on the card.
Conclusion
When LM Studio won’t use your GPU, resist the urge to reinstall everything. Check the GPU Offload slider first and set it to Max. Then confirm the engine that matches your hardware is installed under Settings, then Runtime, and that your graphics driver is current. That order fixes the large majority of cases in minutes. You’ll know it worked when your GPU spikes in Task Manager and your tokens-per-second climbs sharply.
Related Post:
What Does the Ti Mean in a GPU? Nvidia’s Ti Suffix Explained
How to Switch From an NVIDIA to an AMD GPU (Clean, Step by Step)
How to Change Thermal Paste on a GPU (Step-by-Step)
Should You Turn On Hardware-Accelerated GPU Scheduling?
Performance Scaling on GPU or Display: Which Should You Pick?
![Fix GPU memory leak [Forge/Fabric]: what it does and how to install it](https://nobelspot.com/wp-content/uploads/2026/08/GPU-Fan-Speed-33.jpg)
