How to Use LM Studio: The Beginner's Guide (Install, First Model, First Chat, and the Local-Server Branch)

CodexFuse Team · 2026-10-29T11:07:51.934Z · 14 min read

Learning how to use LM Studio is the simplest way to run a real AI model on your own computer instead of sending every question to someone else's server. LM Studio is a free desktop app with a point-and-click interface: you browse its catalogue of open models, download one that fits your RAM, load it into memory, and chat with it completely offline. This LM Studio tutorial takes you from zero to your first private AI conversation in about thirty minutes: what LM Studio is and what it costs, the RAM your computer needs, how to install LM Studio on Windows, Mac, and Linux, how to pick your first model, the one step where beginners stall (downloading a model is not loading it), presets, the local server branch, and how it compares with the Ollama beginner guide if you would rather work in a terminal.

How do you use LM Studio?

You install the app, download one AI model from its Discover tab, load that model into memory, and start chatting in its built-in window. Everything after the download works offline, and local use is free. The five-minute version of the whole process:

  1. Install LM Studio from lmstudio.ai for your system (Windows x64 or ARM, macOS on Apple Silicon, or Linux). Windows x64 machines need a CPU with AVX2 support.
  2. Open the Discover tab and search for a beginner-friendly model such as Qwen or Gemma. Pick a file whose size fits your RAM — LM Studio marks the files your machine can comfortably run.
  3. Click Download and wait. Downloading copies the file to your disk. It does not load it — chat cannot start until the next step.
  4. Open the Chat tab, select the model, and click Load Model. Once it finishes loading into memory, type your first message and chat normally.
  5. (Optional) Open the Developer tab and start the local server if you want other apps to use your model like an API at localhost:1234.

What is LM Studio, exactly?

LM Studio is a desktop application from Element Labs, Inc. that downloads and runs open-weight large language models locally on Windows, macOS (Apple Silicon), and Linux (Nerd Level Tech's LM Studio guide, March 2026). It has three parts a beginner should know: the Discover tab where you browse and download models from the Hugging Face ecosystem, the Chat tab where you talk to a loaded model, and the Developer tab where you run a local OpenAI-compatible server (IBM's Granite guide for LM Studio).

The desktop app is free for local use. That line is worth stating plainly, because LM Studio is not open source — it is a proprietary app, though its lms command-line tool is MIT-licensed on GitHub and the inference engines it runs (llama.cpp and MLX) are open source (Logeshwaran.org's 0.4.26 install guide, checked October 2026). You pay only if you buy cloud features: the Bionic+ plan ($20/month) and Pro plan ($100/month) cover cloud-hosted models and web search inside Bionic, the separate agent app Element Labs launched in July 2026 (Tech-Insider's October 2026 comparison; Sequenced.ai's LM Studio pricing table, consulted October 28, 2026). Running models on your own computer — the desktop app, the lms CLI, and the llmster headless daemon — costs nothing, with no account and no trial period (Logeshwaran.org).

If you would rather not run models on your own hardware at all, the hosted lanes are different products: read the DeepSeek beginner guide for a free hosted model, or the ChatGPT beginner guide for the most familiar hosted chatbot. LM Studio's lane is yours alone: your machine, your models, your chats.

What does your computer need to run LM Studio?

LM Studio's RAM requirements matter more than anything else, because the model you pick is decided by the memory you have — not by which model sounds most impressive. The official-feeling numbers from the guides agree: 8 GB of RAM is the minimum for a comfortable experience, 16 GB is where it gets pleasant, and basic 7B-parameter models want 4 GB or more of GPU VRAM (Nerd Level Tech, March 2026). Logeshwaran.org's install guide (checked October 2026) recommends 16 GB of RAM and 4 GB of VRAM outright.

The operating-system requirements are firm, so check them before downloading a 600 MB installer. Windows runs on x64 CPUs with AVX2 or on ARM64 — a Windows x64 machine without AVX2 cannot run the app at all. macOS requires Apple Silicon (M1 or newer) with macOS 14 or newer; Intel Macs are not supported. Linux needs x64 with AVX2 or aarch64, delivered as an AppImage (all from Logeshwaran.org's install guide, checked October 2026).

Here is the original sizing table I worked out for this guide. The file sizes are computed estimates for typical 4-bit (Q4_K_M) quantizations — about 0.6 GB per billion parameters — plus 1–2 GB for the context window and 3–4 GB of headroom for the operating system:

Model sizeEstimated file size (Q4)Comfortable minimum RAMNotes
1–3B parameters1–2 GB8 GBRuns on almost any modern laptop, including CPU-only machines.
7–8B parameters4–5.5 GB16 GB, or a GPU with 6+ GB VRAMThe sweet spot for a first real model; 4 GB+ VRAM if you use a GPU (Nerd Level Tech).
13–14B parameters8–9 GB24+ GB, or a GPU with 12+ GB VRAMOnly if your machine genuinely has the memory to spare.
70B parameters40+ GB64+ GB workstation RAMNot a beginner target — included so you stop wondering.

The practical rule: pick a 1–3B model on an 8 GB laptop and a 7–8B model on a 16 GB machine, and save the big models for later. The first-model pick belongs in the first five minutes, not in an appendix — everything after it depends on the choice.

How do you install LM Studio?

Download only from the official site, lmstudio.ai. The app is signed and downloaded from one domain (Logeshwaran.org's install guide). As of version 0.4.26 (released October 13, 2026, per LM Studio's own changelog), the installers weigh 619 MB for Windows x64, 485 MB for Windows ARM, 571 MB for Mac, 1.02 GB for the Linux x64 AppImage, and 1.29 GB for Linux ARM64 (Logeshwaran.org, checked October 24, 2026). Models are extra — 2 to 20 GB each, downloaded later.

The steps are the same on every system: download the installer for your OS, run it, and open the app. On Apple Silicon Macs you can also install with brew install --cask lm-studio (Codersera's complete 2026 guide). On Linux, prefer the AppImage or .deb; if an AppImage refuses to start, install FUSE 2 first — a fix the project's own changelog documents for Linux systems.

How do you download and load your first LM Studio model?

Open the Discover tab — the magnifying-glass icon on the left sidebar — and search for a model family such as Qwen, Gemma, Llama, DeepSeek, or Mistral. For each model you will see several files: these are different quantizations (compressed versions) of the same model. A beginner-safe first pick is a small instruction-tuned model such as Qwen2.5-1.5B-Instruct or Llama 3.2 3B (the Cognativ setup guide, updated 3 days ago, walks through Qwen2.5-1.5B-Instruct-GGUF as its worked example). Lower quantizations like Q4_K_M are smaller and faster; higher ones like Q8_0 are better quality but much larger (ExpansiveLabs' deep dive). LM Studio flags the files your hardware can comfortably run — the step-by-step "Day out with LM Studio" guide describes a green rocket icon marking files that fit your machine.

Now the engine pick, which trips up beginners on Apple hardware: on Apple Silicon, prefer MLX model files over GGUF — MLX runs natively on the Mac's unified memory. On Windows and Linux, GGUF files are the pick, with the CUDA llama.cpp runtime on NVIDIA GPUs, the Vulkan llama.cpp runtime on AMD or Intel GPUs, and the CPU llama.cpp runtime if you have no dedicated GPU (the "Day out with LM Studio" step-by-step guide; Codersera's guide recommends MLX before GGUF on Apple Silicon).

Click Download and wait for the file to finish. Then comes the step where beginners actually stall: downloading is not loading. A downloaded model sits on your disk doing nothing. Open the Chat tab, click the model selector at the top ("Select a model to load"), choose your downloaded model, review the load settings — start with a 2,048- or 4,096-token context length, since longer context eats more memory, and set GPU offload to Max on a discrete GPU — then click Load Model (Cognativ's guide; the "Day out with LM Studio" hardware steps). Only when the progress bar finishes is the model actually in memory. Now type your first message and chat.

How do you have your first chat — and what are LM Studio presets?

The Chat tab is your conversation window: type, press Enter, and the answer is generated entirely on your machine. One habit to build immediately: try one offline test. Disconnect your Wi-Fi and ask the model a question. It answers — that is the whole point of local AI, and the moment the privacy promise stops being abstract.

Presets are reusable saved configurations. Instead of re-choosing your model, context length, GPU offload, and inference settings every time you open the app, you save them as a preset and switch with one click — one preset for quick chats with a small model, another for long-document work with a larger context window. Beginners can ignore them for the first day and discover them the moment reconfiguring gets annoying.

LM Studio can also chat with your own documents through retrieval-augmented generation (RAG): attach PDFs or Word files and ask questions about them (Nerd Level Tech's guide; IBM's Granite guide). It is the same privacy deal as chat — the documents never leave your machine.

How do you turn LM Studio into a local API server?

Open the Developer tab and start the server. That is the whole procedure — and it is why developers who began as chat beginners keep the app installed. The server runs on localhost:1234 and speaks the OpenAI API format, so tools built for OpenAI-style endpoints work with your local model: /v1/chat/completions, /v1/models, and in recent versions tool-calling endpoints too (LobeHub's LM Studio provider docs, updated 11 hours ago; IBM's Granite guide).

Two settings matter on the first run. First, enable CORS (Cross-Origin Resource Sharing) in the server settings if a browser-based app like LobeHub needs to reach the server — without it, the browser blocks the request (LobeHub docs). Second, know that the server only serves models you have loaded: load the model in the Developer tab's model loader (Ctrl/Cmd + L), then point your app at http://localhost:1234/v1 as if it were any OpenAI-compatible endpoint.

Practical uses for the local server: connecting coding assistants (such as the Continue extension in VS Code) to a local model for offline autocomplete, running your own scripts against the model with curl or Python, and prototyping app ideas against a model that costs you nothing per token. For headless machines — a server, a CI runner, a GPU rig with no monitor — the lms command-line tool and the llmster daemon run LM Studio's engine with no GUI at all, a mode added in version 0.4.0 (PromptQuorum's LM Studio review, October 2026).

LM Studio vs Ollama: which one should a beginner use?

If you want a graphical app and never want to see a terminal, choose LM Studio. If you are comfortable with the command line and want models inside scripts, agents, and automation, how Ollama works is the better read. They solve the same problem — running open models locally — with different surfaces: LM Studio is a full desktop app with chat, model browsing, and a local server; Ollama is a daemon plus CLI (and optional cloud models) that lives in your terminal (Tech-Insider's $0-to-$500 comparison, October 2026).

The honest differences beyond the surface: LM Studio is proprietary software (free for local use; paid plans only touch cloud features), while Ollama is MIT-licensed open source. LM Studio supports GGUF and MLX model files; Ollama uses GGUF and its own model library. Both offer OpenAI-compatible local APIs. Beginners who like buttons will be chatting in LM Studio within half an hour; beginners who live in a terminal will find Ollama's ollama run flow faster. Neither is wrong — pick the surface you will actually open.

Is LM Studio private and safe?

For local models, yes — and this is checkable, not just a claim. Chats, documents, and model files stay on your machine; after the download, everything works with the network cable unplugged. LM Studio's June 2026 privacy policy states the app has no telemetry or user-specific tracking; the update check sends only your app version and operating system, and model searches send anonymized queries (Logeshwaran.org's install guide, checked October 2026).

The real risk is not the app — it is the model file. A model from an unknown publisher can behave unpredictably, and some model licenses restrict commercial use, so stick to known publishers and read a model's license before using it commercially (Logeshwaran.org). And one general rule of AI hygiene still applies: never feed passwords, API keys, or ID numbers to any AI tool, local or not.

Sources

Frequently Asked Questions

Is LM Studio really free?

Yes, for local use. The desktop app, the lms command-line tool, and the llmster headless daemon are all free downloads with no trial period and no account required. The paid plans — Bionic+ at $20/month and Pro at $100/month — only cover cloud-hosted models and web search inside Bionic, the separate agent app. Running models on your own computer costs nothing (Logeshwaran.org; Tech-Insider, both October 2026).

How much RAM do I need for LM Studio?

8 GB is the minimum for a comfortable experience and 16 GB is where it gets pleasant (Nerd Level Tech). Match the model to the memory: 1–3B models fit 8 GB laptops, 7–8B models want 16 GB of RAM or a GPU with 4+ GB of VRAM, and 70B models belong on 64+ GB workstations. See the computed sizing table above for the full breakdown.

Does LM Studio work on Windows, Mac, and Linux?

Yes, all three, with firm requirements. Windows needs x64 with AVX2 or ARM64. Mac needs Apple Silicon (M1 or newer) with macOS 14 or newer — Intel Macs are not supported. Linux needs x64 with AVX2 or aarch64, delivered as an AppImage (Logeshwaran.org's install guide, checked October 2026).

What is the difference between downloading and loading a model?

Downloading copies the model file to your disk; loading moves it into RAM or VRAM so the app can actually run it. A downloaded-but-not-loaded model cannot chat — you must open the Chat tab, select the model, and click Load Model first. This is the step where most beginners stall (Cognativ's setup guide).

Can I use LM Studio with VS Code or my own code?

Yes. Open the Developer tab and start the local server — it runs at localhost:1234 with OpenAI-compatible endpoints, so coding assistants like the Continue extension and your own scripts can use your local model. Enable CORS if a browser-based app needs access, and load the model before making requests (LobeHub docs; IBM's Granite guide).