A MacBook is a surprisingly good place to run AI privately. The same unified memory that makes Apple Silicon fast also lets a laptop hold and run real language models with no cloud, no account, and no data leaving the machine. This is a step-by-step tutorial: install a runtime, download a model that fits, and chat fully offline. It also covers what most guides skip: why the fanless Air behaves differently under sustained load than a MacBook Pro, and what that means for a long session.
Want the short version? Jump to the summary table. Want the same private AI when your MacBook isn't with you? PocketLLM is designed to run small models on your iPhone fully on-device — coming soon, join the launch list.
Install LM Studio or Ollama, download Llama 3.2 3B (~2 GB), and chat — about ten minutes start to finish. On an 8 GB MacBook stick to 3B models; on 16 GB you can run a 7B like Qwen 2.5 7B. The fanless MacBook Air handles this silently, though sustained non-stop generation will warm it enough for macOS to reduce clocks, since it has no fan to shed the heat. Once the model is local, disconnect Wi-Fi and it keeps working — fully private, fully offline.
PocketLLM is launching soon. Private, on-device AI, starting on iPhone and iPad with more platforms planned. No account, no tracking, no cloud. Join the launch list and be first in.
What you need before you start
- An Apple Silicon MacBook. Any Air or Pro with Apple Silicon works; the only thing that matters is how much unified memory it has.
- Enough free disk and RAM. A 3B model needs ~2 GB on disk and ~4 GB of memory; a 7B needs ~4.5 GB on disk and ~8 GB of memory.
- One internet connection — once. You need the network only to download the runtime and model. After that you can stay offline forever.
Not sure which MacBook you have or want? Check the memory math in our best Mac for local LLMs buyer's guide first.
The steps
1. Install a runtime
Download LM Studio (no terminal, friendliest) or Ollama (one command, developer-friendly). Both bundle the inference engine and a model downloader. We walk through LM Studio's interface in our LM Studio explainer; if you're weighing the options, see Ollama vs LM Studio vs PocketLLM. Installation is a normal Mac app drag-to-Applications or a single install command.
2. Download a model that fits your memory
In LM Studio, search the model catalog; it shows which models your Mac has enough RAM for. On an 8 GB MacBook, pick Llama 3.2 3B. On 16 GB, you can also pick Qwen 2.5 7B for better quality. In Ollama, run a single pull command for the same model. The download is a one-time ~2 GB (3B) or ~4.5 GB (7B).
3. Start chatting
Open a new chat and type. The first reply comes after a brief pause while the weights load from disk into memory, and subsequent replies skip that step because the model is already resident.
4. Verify it's truly offline
Turn off Wi-Fi mid-conversation. The model keeps generating with no interruption. That's the proof that everything is on-device: your prompts and responses never touched a server, and they never will.
What to run, by MacBook memory
| MacBook RAM | Best model | Size (Q4) | Notes |
|---|---|---|---|
| 8 GB | Llama 3.2 3B | 2.0 GB | Don't attempt 7B — it swaps |
| 16 GB | Qwen 2.5 7B | 4.5 GB | Comfortable, room for other apps |
| 24–32 GB | Mistral Nemo 12B | 7.5 GB | Pro territory; long-context work |
The fanless MacBook Air, thermally tested
The MacBook Air has no fan, which raises an obvious question: does it cook itself running an LLM? The honest answer is that it depends entirely on the shape of your usage. Chat is bursty: the chip works for a few seconds per reply and idles while you read, which a passively cooled chassis handles easily. Sustained generation is a different workload, and that is where a fanless design eventually reduces clocks. The way to slow down was a deliberately abusive test: generating thousands of tokens non-stop for many minutes, at which point the chassis warmed and speed dipped slightly as the chip protected itself. For the way people actually use a chat assistant, the Air is silent, cool enough, and completely fine. The MacBook Pro, with active cooling, holds full speed even through those marathon generations.
What it does to your battery
The pleasant surprise is the battery. Apple Silicon only draws power during the seconds it is actually generating tokens, not while you read or type, so chat-shaped use has a very different power profile from a sustained benchmark loop. Non-stop generation pulls considerably more, but that is not how anyone uses a chat assistant. We have not published battery measurements of our own and are not going to estimate one here.
Why this is private by design
There's no privacy policy to trust here, just physics. The weights live on your SSD, the math runs on your chip, and the only network traffic is the one-time download. Disconnect and nothing changes. That architectural privacy — no accounts, no telemetry, nothing to leak — is the same principle PocketLLM follows on the phone. If your MacBook can do this, your iPhone can do a scaled-down version of the same thing.
The quick answer
Install LM Studio or Ollama, download Llama 3.2 3B if you have 8 GB or Qwen 2.5 7B if you have 16 GB, and you're running private AI on your MacBook in about ten minutes. The fanless Air handles it silently and sips battery; the Pro holds full speed on long jobs. Once the model is local, you never need the internet again.
Want the same offline AI when your MacBook is at home? PocketLLM is designed to run small local models on your iPhone, fully on-device with zero telemetry. Coming soon — join the launch list.
Frequently asked questions
How do I run a local LLM on a MacBook?
Install a runtime like LM Studio or Ollama, download a model that fits your memory, and start chatting. The whole process takes about ten minutes on a MacBook Air or Pro. Llama 3.2 3B is roughly a 2 GB download at Q4, so the wait is mostly your connection speed, and once the model is local you can disconnect from the internet and keep using it.
Can a fanless MacBook Air run a local LLM without overheating?
Generally yes, with a caveat about sustained load. A 3B model at Q4 is a small enough working set that ordinary back-and-forth chat only loads the chip in short bursts, between which it idles while you read and type. Because the Air is fanless it relies on the chassis to shed heat, so a very long continuous generation can warm it and cause macOS to reduce clocks. That is a property of passive cooling rather than of local AI specifically.
How much battery does a local LLM use on a MacBook?
Less than you would expect. Apple Silicon runs inference on an efficient integrated GPU and Neural Engine, so the model only draws power during the seconds it is actually generating, not while you read or type. The practical effect is that chat-shaped use draws power in short bursts rather than continuously. Sustained, non-stop generation drains faster, but everyday use is gentle on the battery.
Which model should I run on a MacBook Air with 8GB of RAM?
On an 8 GB MacBook Air, run a 3B model like Llama 3.2 3B, which fits in about 2 GB at Q4. Do not try to run a 7B model on 8 GB — it technically loads but the system swaps and slows to a crawl. If you have a 16 GB Air, you can step up to a 7B model like Qwen 2.5 7B comfortably.
Can I run the same local AI on my iPhone as on my MacBook?
For small models, yes. A 3B model like Llama 3.2 3B that runs on a MacBook Air also runs on a modern iPhone, because the phone has enough unified memory. Larger 7B-plus models need the MacBook's bigger memory. PocketLLM is designed to run the phone-friendly models fully on-device on iPhone with zero telemetry (coming soon), so you can keep the same private, offline setup when you leave your MacBook at home.