Pre-check: open Task Manager → Performance to see RAM and GPU. Then run this in Command Prompt to see GPU name, driver version and VRAM. Also check free disk space.
nvidia-smi
Do you see your GPU name and "Driver Version"?
Download the Windows installer from lmstudio.ai and install. x64 CPUs need AVX2 (most CPUs since ~2013).
LM Studio opened?
In Discover (🔍) search "Llama 3.2 3B" or "Qwen" and download a Q4_K_M GGUF (~2 GB). 3B–4B models suit 4 GB VRAM.
Download finished?
In Chat, pick the model at the top to load it. In load settings, push "GPU offload" as high as fits.
Model loaded and answering?
Pre-check: Task Manager → Performance → GPU shows "AMD Radeon …" and dedicated GPU memory (VRAM). Update the driver with AMD Software: Adrenalin (amd.com).
Can you see the GPU name and VRAM?
Install from lmstudio.ai. AMD Radeon GPUs run through the Vulkan runtime.
App opened?
Download a small model (3B–4B, Q4_K_M) in Discover.
Downloaded?
Make sure the Vulkan runtime is selected (Settings → Runtime), then load the model in Chat.
Getting replies?
Pre-check: Apple menu → About This Mac. Note the Chip (Apple M-series or Intel), Memory (RAM) and macOS version.
Noted chip, memory and macOS version?
Apple Silicon (M1–M4) with macOS 14+ only. Download the .dmg from lmstudio.ai and drag to Applications. Intel Macs are not supported — use the Ollama or GPT4All guide.
App opened?
Download a model in Discover — MLX or GGUF both work on Mac. On 8 GB Macs stay with 1B–3B.
Downloaded?
Load the model in Chat and ask something.
Got a reply?
Pre-check: check RAM, disk and GPU in a terminal.
free -h
df -h ~
lspci | grep -i -E "vga|3d"
See RAM, free disk and GPU name?
Download the AppImage from lmstudio.ai (Ubuntu 20.04+). Make it executable and run it.
Download a small GGUF model in Discover and load it in Chat.
Getting replies?
Pre-check: see your RAM (Task Manager / About This Mac / free -h). Small tools run without a GPU — just slower.
Know how much RAM you have?
Install as in your OS tab. Without a GPU: set GPU offload to 0 and use 1B–3B models. Under 8 GB RAM, Ollama or GPT4All is lighter.
Small model answering?
Sizes are approximate and can change on the official page.
Errors & fixes
"Failed to load model" / "model loading aborted"
Model too big for your RAM/VRAM. Lower GPU offload, reduce context length, or use a smaller (Q4, 3B) model.
Windows: "Your CPU does not support AVX2"
Older CPU — LM Studio will not run. Use Ollama, GPT4All or KoboldCpp (oldpc build).
Linux AppImage: "dlopen(): error loading libfuse.so.2"
Ubuntu 22.04: sudo apt install libfuse2 · Ubuntu 24.04: sudo apt install libfuse2t64. Then chmod +x and run.
"Windows protected your PC" (SmartScreen)
Windows shows this for new apps. Continue only if you downloaded from the official link above: More info → Run anyway.
Download stuck / very slow / "connection reset"
Network or DNS issue. Restart the router, try a mobile hotspot, or set DNS to 1.1.1.1 / 8.8.8.8. Re-running the same command usually resumes the download.
⚠️ MCP servers run with your permissions and can read data you give them; only connect servers you trust. LM Studio asks you to confirm each tool call—keep that on.
Stars, license and last update come from GitHub (17 Nov 2026). Commands are copied from official docs and are not tested on a real machine yet. Access tags are our summary — always check the source. Skills for every AI →