The fastest way to run a local AI on a Mac without touching the command line is LM Studio: a free desktop app with a ChatGPT-style chat window, a built-in model browser, and automatic Apple Silicon (Metal) acceleration. Download it, pick a model that fits your RAM, and start chatting in about five minutes.
This guide walks you through the whole flow on an Apple Silicon Mac (M1 through M4). It covers which model to pick by RAM, one common trap that makes beginners think local AI is broken, and how LM Studio compares to the Ollama command-line route.
What Is LM Studio?
LM Studio is a free desktop app for macOS, Windows, and Linux that runs open-weight language models entirely on your own machine. Nothing leaves your Mac — your chats stay private and work fully offline once a model is downloaded.
It bundles three things beginners usually have to wire up by hand:
- A chat interface that looks and feels like ChatGPT
- A model browser that downloads models with one click
- A local inference engine with Metal GPU acceleration built in
You do not open a terminal, edit a config file, or install anything else. That is the whole pitch.
Why Skip the Terminal?
Most local-AI tutorials start with Ollama and command-line installs. That route is powerful and scriptable, but it is a wall for anyone who just wants a private chatbot.
A graphical app removes three friction points at once: you see every available model with its size and RAM cost, you click to download instead of memorizing tags, and you chat in a normal window instead of a text prompt. For first-timers, that difference decides whether local AI sticks or gets abandoned.
If you later want automation or an API, the Ollama setup guide covers the command-line path. Many people run both.
Step 1: Install LM Studio
Go to lmstudio.ai/download and grab the Apple Silicon build.
1. Open the downloaded .dmg file
2. Drag LM Studio into your Applications folder
3. Launch it (right-click → Open the first time if macOS warns about an unidentified developer)
The app detects your chip and available memory on first launch, so its suggestions already account for your hardware.
Step 2: Pick a Model for Your RAM
This is the step that makes or breaks the experience. A model that is too big for your RAM will swap to disk and crawl. The table below maps unified memory to a safe model size on Apple Silicon.
| Your RAM | Safe model size | Good first pick |
|---|---|---|
| 8 GB | 3–4B | Qwen3.5 4B |
| 16 GB | 7–9B | Qwen3.5 9B |
| 24–32 GB | 14–32B | Qwen3.5 35B-A3B |
| 64 GB+ | 70B class | Larger frontier-class open models |
A practical rule on Apple Silicon: keep the model's download size under about 70% of your total RAM, leaving headroom for macOS and your apps. On a 16 GB Mac that means staying near 7–9B parameters at 4-bit quantization.
Not sure what your Mac can handle? The best local LLM for MacBook guide and the model finder wizard rank picks for your exact chip and memory.
Step 3: Download a Model
In LM Studio, open the search / discover tab (the magnifying glass) and type a model name — for example qwen3.5.
1. Find the size that matches your RAM (e.g. the 9B for a 16 GB Mac)
2. Click Download — LM Studio shows the exact disk size before you commit
3. Wait for it to finish (a 9B model is roughly 6 GB)
Once downloaded, the model lives on your Mac permanently and never needs the internet again.
Step 4: Chat
Open the chat tab, select your model from the dropdown at the top, and type. The first message loads the model into memory (a few seconds), then responses stream back token by token like any cloud chatbot.
That is it. You now have a private AI assistant running on your Mac.
Watch Out: The "Reasoning Leak" Trap
Here is the trap that makes beginners think local models are useless. Some models are reasoning models that, with the wrong settings, dump their internal thinking — long monologues, raw JSON, self-talk — straight into the chat. You ask "how's it going?" and get a 200-word internal trace plus a stray code block.
That is not a broken app. It is a chatty reasoning model shown without its thinking pane. Two fixes:
- Switch model. For casual chat, pick a clean instruct model like Qwen3.5 4B/9B instead of a reasoning-heavy one.
- Disable reasoning. In the chat settings, turn off "reasoning" / "show thinking," and add a system prompt like "Answer concisely. No visible reasoning. No JSON."
For a first model, a plain instruct model gives the cleanest ChatGPT-like experience.
LM Studio vs Ollama: Which Should You Use?
| LM Studio | Ollama | |
|---|---|---|
| Interface | Graphical, ChatGPT-style | Command line |
| Model install | One click | ollama run <model> |
| Best for | Beginners, daily chat | Developers, scripting, APIs |
| Setup time | ~5 min | ~5 min + terminal comfort |
| Cost | Free | Free |
Neither is "better" — they target different users. Start with LM Studio for a no-terminal chat experience. Move to Ollama when you want to script, build apps, or expose a local API. The run AI offline guide shows how to combine them.
Frequently Asked Questions
Is LM Studio free?
Yes. LM Studio is free for personal use and runs entirely on your own hardware. There is no subscription and no per-message cost, because the model runs locally rather than calling a paid API.
Does running a local LLM need the internet?
Only to download the app and models. After that, everything runs offline. Your prompts and responses never leave your Mac, which is the main privacy advantage over cloud chatbots.
Which model should a 16 GB Mac start with?
Qwen3.5 9B is a strong all-round pick for a 16 GB Apple Silicon Mac — good answer quality at a download size near 6 GB. If you want maximum speed, drop to Qwen3.5 4B (around 3 GB). Both leave comfortable RAM headroom for macOS.
Why are some local models so verbose?
They are reasoning models leaking their internal thinking into the chat. Switch to a plain instruct model, or disable the reasoning option in chat settings and add a system prompt asking for short answers without visible reasoning.
Is LM Studio faster than Ollama on Mac?
Speed is roughly the same — both use Metal GPU acceleration on Apple Silicon and similar inference backends. The real difference is the interface, not raw throughput. Token speeds depend on your chip and the model size, and any figures here are estimates, not measured benchmarks.
Can I use LM Studio and Ollama together?
Yes. They do not conflict. Many people use LM Studio for everyday chat and Ollama for scripting or running a background API server. Models downloaded in one are separate from the other.
Next Steps
You now have a private, offline AI on your Mac with no terminal required. From here:
- Browse the model finder to match a model to your exact chip and RAM
- Compare picks in the best LLM for MacBook guide
- When you want automation, follow the Ollama setup guide
Where to Buy for Local AI
best configsRuns 30B models with headroom; active cooling sustains long inference without throttling.
Loads 70B models locally — the most capable AI laptop config.
ModelFit may earn a commission on purchases made through these links, at no extra cost to you. Recommendations are based on local-AI performance, not commissions.
The weekly local-AI refresh
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
Have questions? Reach out on X/Twitter