Updated September 2026 for LM Studio 0.4.x and the latest open models. First published August 2025.
Running an AI model on your own computer is no longer a hobby for geeks. In 2026, a recent laptop with 16 GB of RAM can run models that rival the cloud assistants of two years ago, fully offline, for free, and without a single prompt leaving your machine. This guide walks you through setting up LM Studio, the easiest way to get started: what hardware you need, which model to pick, how to install it step by step, and how to use it as a local API. The steps are almost identical on Windows, macOS and Linux.
📕 Once your local AI is running, you need to know what to ask it. My book L’IA générative pour tous (in French) collects 150+ ready-to-use prompts that also work offline: see it on Amazon.
Why run AI locally?
- Privacy: your documents and prompts stay on your machine. Nothing is sent to a remote server, which makes local AI a strong option for sensitive data (HR, legal, health, finance).
- Offline: it works on a plane, on a train, or on a locked-down network.
- No subscription: LM Studio is free, including for work use. Your only cost is the hardware you already own.
- Control: you choose the model, its version and its settings, and nobody changes them behind your back.
The limits you should know about
Local AI is powerful, but it is not magic. Set realistic expectations before you start:
- Memory is the bottleneck: the model must fit in your RAM (or your GPU’s VRAM). Bigger models are smarter but need more memory.
- Speed depends on your hardware: without a decent GPU or an Apple Silicon Mac, answers come more slowly than with ChatGPT or Copilot.
- The very best models stay in the cloud: the largest frontier models need data-center hardware. Locally, you get very good quality for everyday tasks, not the absolute top.
- Energy and disk space: a model weighs from 2 GB to more than 60 GB, and heavy use drains a laptop battery quickly.
Why LM Studio?
LM Studio is a free desktop application with a clean graphical interface. You can search, download and chat with open models without touching the command line. Its CLI (lms) and SDKs are open source (MIT). The current version, LM Studio 0.4.25 (September 2026), brings:
- a built-in model catalog connected to Hugging Face, with GGUF and MLX (Apple Silicon) formats;
- multimodal models: you can drop an image or a PDF into the chat;
- a local server compatible with the OpenAI API, now with image support on
/v1/chat/completionsand/v1/responses; - MCP support, to connect your local model to tools;
- speculative decoding for faster answers, and a new faster engine on Apple Silicon M3 and later;
- LM Link, to use a model running on one computer from another device (including the « Locally » app on iPhone and iPad);
- chat export to PDF and Mermaid diagram rendering.
Note: LM Studio is built for running models (inference), not for training or fine-tuning them. For fine-tuning, you need dedicated tools such as Unsloth.
System requirements
- macOS: macOS 14 or newer on Apple Silicon (M1 to M4 and later). This is the most comfortable platform: RAM is shared with the GPU.
- Windows: x64 (CPU with AVX2) or ARM (Snapdragon X). A GPU with at least 4 GB of dedicated VRAM is recommended.
- Linux: Ubuntu 20.04 or newer, x64 with AVX2 (AppImage).
- Memory: 16 GB of RAM recommended. 8 GB works with small models.
- Disk: an SSD with at least 50 GB free if you want to try several models.
Which model should you choose in 2026?
The right model depends first on your memory. Here is a simple starting point, using the default 4-bit quantized versions (Q4_K_M or equivalent) offered by LM Studio:
| Your memory (RAM or VRAM) | Model to start with | Best for |
|---|---|---|
| 8 GB | Gemma 4 E4B (~6 GB) or Phi-4-mini (~2.5 GB) | Emails, summaries, rewording, quick questions |
| 16 GB | gpt-oss-20b (~12 GB) | Reasoning, structured writing, tool use |
| 24 GB | Gemma 4 26B A4B (~16 GB) or Qwen3.8 27B (~24 GB) | Analysis, documents and images, coding |
| 64 GB and more | gpt-oss-120b (~65 GB) | Near-cloud quality for demanding work |
My advice: start small. A fast, small model that answers in two seconds is far more pleasant to use than a big one that takes a minute. Move up only if the quality is not enough for your use case. Gemma 4 and Qwen3.8 also read images, which is handy for screenshots and scanned documents.
Step-by-step installation
Count 10 to 20 minutes, most of it spent downloading the model.
1. Download and install LM Studio
- Go to the official site: lmstudio.ai.
- Download the installer for your system (.exe for Windows, .dmg for macOS, AppImage for Linux).
- Install it and launch it once. The
lmscommand-line tool is bundled: after this first launch, typelmsin a terminal to check it works.
2. Download your first model
- Open the Discover tab (magnifying glass icon).
- Search for the model that matches your memory (see the table above), for example gpt-oss-20b or Gemma 4.
- LM Studio shows whether the model fits your machine. Keep the recommended quantization and click Download.
From the terminal, the same thing takes one line: lms get openai/gpt-oss-20b. The LM Studio model catalog lists all available models with their size.
3. Chat with the model
- Go to the Chat tab and select your model at the top of the window.
- Type your first prompt, for example: « Summarize this document in five bullet points » and drop in a PDF.
- For reasoning models such as gpt-oss, you can set the reasoning effort (low, medium, high): low for quick answers, high for complex questions.
4. Optional: use it as a local API
LM Studio can act as a local server compatible with the OpenAI API. Any tool that supports a custom OpenAI endpoint can then use your local model.
- In the app: open the Developer tab and switch the server to Running.
- Or in a terminal:
lms server start, thenlms load openai/gpt-oss-20b. - The server listens on
http://localhost:1234/v1. Openhttp://localhost:1234/v1/modelsin a browser to see your models. - Other useful commands:
lms ls(models on disk),lms ps(loaded models),lms chat(chat in the terminal).
Tips for a better experience
- Slow answers? Choose a smaller model or a lighter quantization, and close other heavy applications.
- GPU not used? Update your graphics drivers, then check the GPU offload setting when loading the model (
lms load --gpu=max). - Long documents? Increase the context length when loading the model, but keep in mind that it uses more memory.
- Keep LM Studio up to date: new versions bring speed gains and support for new models every few weeks.
Alternatives to LM Studio
- Ollama: command-line first, very popular with developers and easy to script.
- Jan: fully open source, simple interface, 100% offline.
- GPT4All: beginner-friendly, with built-in chat over your local documents.
Local AI or Copilot in the enterprise?
As a CIO, I see local AI as a complement, not a replacement. It is ideal for testing, for sensitive data, and for people who want full control. For company-wide use, with access to your emails, files and Teams meetings, a managed tool such as Microsoft 365 Copilot is usually easier to govern. If you are weighing these options, my guide on which model to pick in Copilot M365 covers the cloud side of the question.
FAQ
Is LM Studio free?
Yes. LM Studio is free for personal and work use. The models you download are open-weight models, each with its own license (often Apache 2.0 or similar).
Can I run a local AI without a GPU?
Yes, on the CPU alone, as long as your processor supports AVX2. Choose a small model (Phi-4-mini or Gemma 4 E2B/E4B): answers will be slower but perfectly usable. On a Mac with Apple Silicon, the integrated GPU is used automatically.
How much RAM do I need?
8 GB is enough to start with small models. 16 GB is the comfortable minimum and lets you run gpt-oss-20b. With 24 GB or more, you can run mid-size models such as Gemma 4 26B or Qwen3.8 27B.
Does my data stay on my computer?
Yes. Once the model is downloaded, the conversation runs entirely on your machine and works without an internet connection. Only optional features such as LM Studio’s cloud models or LM Link use the network.
LM Studio or Ollama?
LM Studio if you want a graphical interface and a visual model catalog. Ollama if you mainly work in the terminal or want to script and automate. Both expose a local API and can be used side by side.
Which prompts work best with a local AI?
The same ones that work with ChatGPT: give the model a role, some context, the expected format and the tone. Start with simple tasks such as summarizing, rewording or extracting information. Smaller local models reward clear, structured prompts even more than cloud models do.
Is local AI safe for company data?
It removes the main risk, since nothing leaves the machine. You still need to secure the device, limit which data you process and document the use. Before rolling AI out more widely, check that your data is ready: my article Your Data Isn’t Ready for Copilot explains what to audit first.
Conclusion
In 2026, setting up a local AI with LM Studio takes less than half an hour, costs nothing, and keeps your data on your machine. Start with a small model that fits your memory, test it on your real tasks, and move up only if you need to. The best way to learn is to try it yourself. Share your setup and your favorite model in the comments!
📕 Going further: use AI well, every day
Running an AI model locally is a great first step. Getting real value from it every day is the next one. In my book L’IA générative pour tous (in French), I share 150+ tested, ready-to-use prompts for ChatGPT, Claude and Copilot, with no jargon and plenty of real-life use cases.
🤝 Considering local or private AI for your business? As a part-time CIO, I help SMEs choose, deploy and govern AI tools (local LLMs, Microsoft 365 Copilot) while keeping their data under control. Let’s work together →
Sources: LM Studio changelog; LM Studio system requirements; LM Studio CLI documentation; Gemma 4 on LM Studio; gpt-oss on LM Studio.
Written by Sylvain Jacquemard, CIO with 25+ years in IT, Microsoft Global Community Initiative Regional Leader (Switzerland) and author of books on generative AI.
Votre commentaire