Install LM Studio to try local models
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
Goal for tonight: install LM Studio, download a modest Q4 GGUF, and get a reply in the Chat tab — with the terminal left closed.
Four moves (before the technical detail)
- Download the installer from lmstudio.ai and run the wizard.
- Open Discover and Download a small Instruct model in Q4_K_M.
- Load the model in Chat and raise the GPU Layers slider in steps.
- Send a prompt and compare the result with your Ollama stack if you already have one.
LM Studio is a desktop app (Windows, macOS, Linux) for downloading local models and talking to them. The buttons below use the labels you see on screen. This guide is for beginners who want a graphical test bench before they lock a model into Ollama for the whole house (Open WebUI, family assistant).
A short glossary (before the requirements)
- LLM (large language model): a model that writes text from your prompt.
- GGUF (GPT-Generated Unified Format): the quantized weight file you download in Discover before you chat.
- VRAM: graphics-card memory. Putting part of the model there speeds generation up (detail in section 3).
- Layers: blocks of the model’s network, which you can split between GPU and CPU.
- Q4: a quantization level. A good size-versus-quality compromise on a home PC. Prefer Q4_K_M on the first try.
- Tokens/s: tokens per second (a rough stand-in for words per second). It tells you how fluid the chat feels.
What you need
- Windows 10/11, macOS (Apple Silicon is the comfortable choice), or 64-bit Linux.
- 10–20 GB free at minimum for a first Q4 model.
- 16 GB RAM or more. 32 GB is comfortable for a 7B with room to spare.
- A GPU is optional and useful: NVIDIA with current drivers, or an iGPU or Apple Silicon for small models.
- Internet for the installer and the Discover download.
For the hardware decision that comes first, choose hardware for home AI and which model for your VRAM keep you from downloading a 13B onto a machine with 8 GB of RAM.
Time
25 minutes the first time (installer, a first 3B–7B Q4 GGUF, and the GPU setting). Later sessions take a few minutes once the model is already cached locally.
Ollama and LM Studio work side by side
Labels as they appear in the app. The right-hand column is the on-screen wording in LM Studio.
| What you need | Ollama | LM Studio | Label in the app |
|---|---|---|---|
| A first AI in 15 minutes, in a terminal | The better fit — see first local AI | Possible, and heavier | — |
| A local API for Open WebUI and scripts | Built in and stable — see the multi-OS Ollama install | A server mode exists | Local Server |
| Trying five different GGUF files in one evening | Repeated ollama pull |
A visual comparison and GPU sliders | Discover / Search |
| Fine control of VRAM and layers | Automatic, based on the GPU | Manual control of the layers | GPU Layers / GPU offload |
| A household that does not want a terminal | Open WebUI on top of Ollama | A friendly built-in conversation | Chat |
Ollama commands, smart-home integrations, and the server stack live in one place: install Ollama on every OS. The Windows guide already covers Ollama’s role, so this page does not repeat it.
1. Download and install from lmstudio.ai
Go to the official site, lmstudio.ai (the main source for this guide). Click Download, pick your system, save the file, and run it. On Windows, the usual installer copies LM Studio into Program Files and creates a shortcut. On Mac, drag the app into Applications. On Linux, follow the format the page offers that day (AppImage or a package).
On first launch, the app asks where to store models. Pick a drive with room to spare. A 7B Q4 is often over 4 GB. If drive C: is tight, point at a second SSD. See SSD storage for models.
Confirm LM Studio is current: Help, then Check for updates, matching the version shown on lmstudio.ai/docs.
2. Find a model in Discover (Q4)
- Open the Discover tab (or Search, if your version says that).
- Type a familiar name:
Llama 3.2 3B Instruct,Mistral 7B Instruct, orQwen2.5 7B Instruct. - Filter or read the quantization line. Prefer Q4_K_M or Q4_0 for a first test. That balance fits a household, and it matches choose a first model.
- Click Download and wait for the bar to finish. Leave LM Studio open during the transfer.
Why Q4? The weights take less disk and less RAM or VRAM than a Q8, with enough quality for chat and household summaries. A 3B Q4 fits a modest PC. A 7B Q4 needs more room. See 32 vs 64 GB of RAM if you are unsure.
If Discover does not list a model, check your connection or try later. You can also import a GGUF you got somewhere else. Discover is still the simplest way to start.
3. Load the model and understand GPU offload
Once the download finishes:
- Go to Chat (or the model-loading screen).
- Select the GGUF you just downloaded.
- LM Studio loads the model weights into memory (the GGUF file on disk moves into RAM, and into VRAM when it can). The first open can take 30 seconds to 2 minutes.
What “GPU offload” means
A whole model does not always fit in VRAM. LM Studio can split the network’s layers:
- some on the GPU (fast);
- the rest in RAM, computed by the CPU (slower).
The GPU Layers slider (also called GPU offload: sending part of the work to the graphics card) sets how many layers run on the GPU. Raising the slider:
- speeds generation up while VRAM has room;
- can crash the app, or show “out of memory,” if you pass what the card can hold.
A careful method: start at 0 layers, raise it by 5 or 10, send a test sentence, and watch. When VRAM is almost full in Task Manager (Windows) or Activity Monitor (Mac), step back one notch. For numbers and common RTX cards, the LM Studio GPU offload guide covers the settings, so this page does not repeat the whole VRAM table.
With no dedicated GPU, leave GPU layers at 0. LM Studio then runs CPU only, the same idea as Ollama on a thin laptop. Slow, and still a useful lesson.
4. A test chat
In the Chat tab, with the model loaded and the layers set, paste this prompt. It asks for French on purpose, so you can see that the model followed the language instruction. After that reply looks right, ask the same thing in English.
Suggested prompt:
Résume en deux phrases ce qu’est LM Studio pour un débutant en IA locale. Réponds uniquement en français.
You should get readable text that mentions an app, local models, and possibly the word GGUF or chat. The speed shown (tokens/s) depends on GPU offload. Above about 10 tokens/s on a 7B, the chat is comfortable enough to read aloud in the living room.
Try a longer prompt — for example, summarizing a made-up school email — to feel the context limit of the model you picked. For a structured family setup, you will come back later to the private family assistant on Ollama or Open WebUI.
What a finished install looks like
At the end of this path, you should see the following:
| Check | A normal result |
|---|---|
| App | LM Studio opens with no error, and the version is visible |
| Model | At least one Q4 GGUF listed as local, or ready in Chat |
| GPU | If an NVIDIA, AMD, or Metal device is available, some layers sit on the GPU and the app stays up |
| French chat | A coherent French reply to the test prompt |
| Disk | The models folder contains a multi-gigabyte file |
If the chat answers in English even though the prompt asked for French, add “Réponds en français” at the start of the message. Some base models follow instructions less closely than Instruct variants.
If the app freezes while loading, lower the GPU layers or pick a smaller 3B Q4, then raise them gradually after a hardware upgrade (hardware selector).
Mistakes that show up often
| Problem | What to do |
|---|---|
| VRAM is full, black screen, or a crash | Lower the GPU layers. Close games or a heavy browser |
| Model marked “corrupted” | Delete the partial file in the LM Studio folder and download it again from Discover |
| Extremely slow | CPU only, or too few GPU layers. Accept it for the test, or aim for a 3B |
| Conflict with Ollama | The two can live on the same PC. Avoid two large models at once on 16 GB of RAM |
| You need an API | Ollama plus Docker Open WebUI is the stack with the most documentation for n8n automations |
Where to go next
- LM Studio GPU offloadFine settings for GPU layers and VRAM.
- Install OllamaThe server engine for the whole house.
- Open WebUIA family interface on top of Ollama.
- Choose a first modelInstruct tags and Q4 quantization.
Optional hardware (after the first chat)
When to use this: only after the install checks out (the test chat works) and the GPU Layers slider tops out low, or generation stays too slow on a 3B Q4. The bottleneck is the PC, not LM Studio.
How: write down the GGUF you tested, the highest GPU layer count that did not crash, and the speed shown (tokens/s). Then open the hardware selector. It lines those signals up with budget, living-room noise, and disk space. Read choose hardware for home AI before any order. The selector does not replace the hardware requirements.
Affiliation: the selector may offer partner links (including Amazon). LocAIhome can earn a commission at no extra cost to you. See the affiliate disclosure. This page does not lock a price or a product listing. Check the current price range on the store (October 2026).
Privacy and limits
LM Studio runs models locally by default for the built-in chat. Read the network settings if you turn on a Local Server that the LAN can reach. Cross-check privacy for local AI data before you paste medical or bank documents. LLMs hallucinate. A local summary is still a draft you should reread.
Recap
You install LM Studio from lmstudio.ai, download a Q4 GGUF through Discover, adjust GPU offload so VRAM and speed stay in balance, and confirm a chat. Keep Ollama for smart-home and API work later. Use LM Studio as a graphical test bench. Fine tuning belongs in LM Studio GPU offload once the first model runs steadily.
Common questions
Does LM Studio replace Ollama in a connected home?
Ollama stays the simpler choice for a stable local API (Open WebUI, n8n, Home Assistant). LM Studio is excellent for comparing GGUF files on screen, setting GPU offload by hand, and chatting with no terminal.
Which Q4 model should I pick the first time in Discover?
Aim for a 7B or 8B instruct model in Q4_K_M if you have 16–32 GB of RAM and a GPU with 6–8 GB of VRAM. Otherwise use a 3B Q4 so the install check does not fill memory.
The Discover download stops. What should I do?
Resume it from the Downloads tab in LM Studio. Delete the partial file if the app offers that, then start again. Check free disk space (often 5–10 GB for a 7B Q4 model).
Why is my GPU missing from the settings?
Update the NVIDIA or AMD drivers, restart LM Studio, and on Windows confirm you are not in a Remote Desktop session with no 3D acceleration. Apple Silicon uses Metal automatically on recent versions.
Sources
Narrow it down with the hardware selector · French version of this section