How to install LM Studio and run your first local model
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
LM Studio lets you download a language model, load it into memory and chat through a normal graphical window. You do not need a terminal for the first conversation. There is no magic, though: a large model still eats RAM, graphics memory and disk space. Check the computer, choose a compact quantized variant, run a short test and only then touch the heavy stuff.
See what LM Studio does without commands
The app combines model search, downloads, memory loading, chat and a local API in one window. That is easier for a beginner than manually juggling files, parameters and commands. You can inspect model variants and download status in the interface.
LM Studio is not one neural network. It is a shell and runtime. Answer quality depends on the selected model, quantization, context length, system prompt and computer. A polished window does not turn a weak model into a genius.
Check whether the computer can handle it
Official requirements recommend 16 GB of RAM. Windows x64 needs an AVX2-capable processor, and 4 GB of dedicated VRAM is a reasonable recommendation. On Apple computers, Apple Silicon and macOS 14 or newer are supported, while Intel Macs are currently out.
Successful installation does not mean every model will run. Model files take gigabytes and need memory when loaded. Leave headroom for the operating system. If swap starts, fans scream and answers arrive one word at a time, the hardware is telling you to stop this nonsense.
Choose the first model without showing off
In Discover, look for an instruct model, a variant prepared for dialogue and instruction following. Search works by name and can also accept a model link. A single model often has several quantized files with labels such as Q4 or Q8.
Quantization reduces file size and memory demand at some cost to fidelity. A 4-bit or better variant that fits available memory is usually a sensible first test. Do not download the fattest file just because the number looks impressive. That is playground logic.
Load the model into memory
After downloading, move to Chat, open the model loader and select the file. Loading allocates memory for weights and runtime parameters. When memory is insufficient, startup may fail or the entire machine may become painfully slow.
Do not max out context and every advanced setting on the first run. Keep reasonable defaults, measure basic speed and change one parameter at a time. Otherwise you get a slideshow and no clue which damn knob caused it.
Test the model with a short conversation
Create a new chat and ask a concrete question with a verifiable answer. Then request a revision or another output format. This reveals whether the model follows instructions, keeps context and responds fast enough on your hardware.
Do not start with confidential documents and do not judge the model by one pretty answer. Try a short summary, text structure, small code fragment or classification task. Local execution is useful, but the model can still produce confident garbage.
Understand what really works offline
After the model and required runtime are downloaded, local chat, document work and the local server can operate without the internet. Conversation data is processed on the device. That is useful for private tests and places with lousy connectivity.
Searching and downloading new models, fetching statistics, obtaining runtimes and checking updates require a connection. Prepare offline mode in advance. Disconnecting before the files exist and then swearing at an empty search is a pretty stupid trick.
Start the local server after the chat works
The Developer tab can start an API server with a switch and no terminal. By default it is available on localhost:1234. Your own application can then call the local model through REST and compatible endpoints.
Keep it on localhost first. The API may have no authentication by default, although a token can be enabled in settings. Before network access, configure a token, firewall and allowed clients. A public unprotected port is not brave. It is a hole.
Verdict: easier interface, same hardware limits
LM Studio is a solid first step into local models without a command line. The sensible order is hardware check, suitable instruct model, memory load, short chat and only then a server or larger context.
The common mistake is treating a friendly interface as the end of technical limits. It removes chores but adds no memory and cannot make every model fast. Match model size to the machine, keep the server closed and do not overcomplicate the first run. Then the whole thing works without needless pain in the ass.
Translation reviewed:
