🔬 LM Studio: the full local AI workstation — hub, chat, server, CLI

What LM Studio actually is

LM Studio is the most complete “local AI workstation” out there. One app, four tools:

1. A model Hub — browse, compare quantizations, and download models (GGUF files) without leaving the app.
2. A chat interface — load any model you have and talk to it, with presets that control sampling.
3. A local server — an OpenAI-compatible API on http://127.0.0.1:1234, so the same apps that point at OpenAI can point at your machine.
4. The lms CLI — every one of those actions available as a command, for scripts and pipelines.

If Ollama is the “3-command” option, LM Studio is the one that also gives you a window to click in — which is exactly what a lot of people actually want.

Step 1 — Install

Windows (installer or winget):

winget install --id ElementLabs.LMStudio

macOS (Homebrew or direct download):

brew install --cask lm-studio

The publisher is Element Labs (the team behind the LLaMA-Factory ecosystem tooling). Launch the app — it’s a normal desktop program, and the first time it opens it asks for basic settings (theme, model folder).

Step 2 — Get a model (the Hub)

Left sidebar → Hub → search for a model — e.g. Qwen3 4B. Each model card shows the available quantizations with their file sizes. Pick Q4_K_M: it’s the sweet spot between quality and size for a first run (~2.5 GB for a 4B model).

Two things worth knowing:

• The Hub downloads GGUF files — the same format Ollama and llama.cpp use, so models are portable between all three. What a GGUF is →
• If a model isn’t in the Hub, the CLI accepts a Hugging Face link directly (below). The Hub and Hugging Face are two doors into the same room.

Step 3 — Chat with it

Left sidebar → My Models → click the model → Chat. Load it in memory (a few seconds for a 4B) and type. The preset selector at the top controls sampling — temperature, top-p, max tokens. You don’t need to touch it to have a conversation.

That’s the GUI part. Now the part that makes it infrastructure.

Step 4 — The local server (port 1234)

LM Studio exposes an OpenAI-compatible API. Two ways to start the server:

• In the app: the Server tab — pick the model and quantization, click Start Server.
• From the terminal:

$ lms server start
$ lms server status
The server is running.

Check it’s alive and see what it serves:

$ curl http://127.0.0.1:1234/v1/models
{
  "data": [
    { "id": "qwen3-4b", "object": "model" }
  ],
  "object": "list"
}

And a real completion, OpenAI format:

$ curl http://127.0.0.1:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-4b","messages":[{"role":"user","content":"Say hi"}]}'

Point any OpenAI SDK at base_url = "http://127.0.0.1:1234/v1" and your code runs on the model in your machine instead of in a datacenter. Why that matters →

The lms CLI (the part developers love)

The lms binary ships with the app and is the fastest way to script anything:

A realistic workflow:

$ lms get qwen/qwen3-4b@q4_k_m      # 1. download (or: lms import path\to\file.gguf)
$ lms chat qwen3-4b                 # 2. talk to it
$ lms server start                  # 3. expose it on :1234

The import flag worth memorizing is --hard-link: on the same volume it makes the existing file available to LM Studio without copying a single byte — a 2.5 GB model import takes under a second.

What to expect from your hardware

Rule of thumb: file size × ~1.5–2 ≈ what you need in RAM/VRAM, context included. Quantization is why this fits at all →

Ollama, llama.cpp or LM Studio?

All three run the same GGUF files on the same llama.cpp engine under the hood. The difference is the surface:

If you’re still choosing, the short answer: LM Studio to explore, Ollama to automate.

Where this leaves you

You have models, a chat, a server, and a CLI. The natural next step is the thing local AI is really for: letting an agent do things on your machine with a local model behind it.