The model file has to fit before the demo matters
Running Llama 3 locally on a Mac means the weights stay on that computer and the answers are computed there, without a chat subscription in the loop. It matters when the prompt contains material you will not paste into a website, or when you want the model available on a plane. It does not matter as a cheaper clone of the largest hosted assistant. A laptop-sized model is a smaller mind. Treat it that way.
This is a step sequence for a Mac you already own. It is not a shopping list and it does not promise a tokens-per-second figure. Those numbers depend on the chip, the memory, and the exact file. The hardware guide is where to think about buying. Stay here if the computer is in front of you.
What you need before the first command
You need a Mac with enough unified memory for the file you intend to open, an admin account, and a sense of whether the prompts are allowed on that disk. Unified memory is shared by the system and the model. If the machine is already swapping when you open a browser and a design app, a large model will not feel local. It will feel stuck.
Pick a small instruct model for the first run. "Llama 3" is a family. The 8B-class instruct files are the usual first fit on machines with 16 GB. Larger files want more memory than a slogan admits. If the app offers a download size, believe the size. A 40 GB file will not become small because the name says efficient.
You also need a way to quit and delete. Before you admire an answer, know which folder the app used. If you cannot delete the model, you do not have a local setup. You have a tenant.
Step 1: Install one runner, not three
Ollama and LM Studio are the two runners that cover almost every first week. Ollama is a small background service and a terminal. LM Studio is an app with a chat window and a model browser. Pick one so you are not debugging two download folders.
Install it from the project's own site, not from a random "Llama 3 DMG" in a forum. On a work Mac, use the install path your admin already allows. If the machine is managed, a local model may be against policy even when the files never leave the disk. Ask before you download weights to a company laptop.
Open the app once and refuse any default that sends prompts to a hosted fallback. The point of this guide is the opposite. If a toggle says "use cloud when the local model fails," turn it off until you understand it. A silent fallback will train you to think you are offline when you are not.
Step 2: Download one instruct model and ask a question you know
In Ollama, the usual first pull is a small Llama 3 instruct tag published by the project. In LM Studio, search the same family and choose a quantization the app says will fit. Then ask something you can grade without research: a sentence you rewrite, a function you already understand, the name of a file you just created.
Read the answer for obedience, not for genius. Did it follow the format you asked for? Did it invent a file that is not on the disk? Local models are confident about your machine and they cannot see it unless the app has a tool you enabled. If you did not enable a tool, an answer that claims it listed your downloads folder is fiction.
The beginner pipeline walks the same first question for people who are not on a Mac. Use it if the menu names in this guide do not match the version you installed. Menus move. The test does not.
Step 3: Prove it is local
Turn off Wi-Fi and ask again. A local model still answers. A wrapper that was calling an API will spin or error. This single check is worth more than the settings screen.
Then look at Activity Monitor while it generates. The memory number should rise by something like the size of the model file, not by a trickle. If memory barely moves and the answer is instant and brilliant, you may still be on a remote model. Check the host name in the app's logs or in Little Snitch or the firewall if you have one. You are looking for an absence of new connections during the reply.
Step 4: Put a real document through it
Paste a page you wrote, not a famous article. Ask for a shorter version that keeps the numbers. Compare the numbers. Local models drop digits and flip conditions, same as hosted ones, sometimes more so because the model is smaller.
If the page is confidential, this is the moment the setup earns its keep. The page did not go to a chat website. It did sit in the app's history. Clear that history, or know where it is stored, before you share the Mac user account. A local model is not a private model if the transcript syncs through iCloud or a crash reporter. Turn those off for the app if the prompts require it.
Step 5: Make an alias you will actually use
A local model you have to rediscover is a model you will not use. In Ollama, a one-line shell alias is enough. In LM Studio, pin the model you tested, not the larger one you have not. Delete the other downloads. Disk space is the silent cost, and a folder of unused weights is how a "free" setup fills the drive.
Write down three lines in a note: the app, the exact model name, and the folder. Future you will not remember the tag. The glossary can remind you what "instruct" and "quantization" mean when the download list gets cute.
What not to expect
Do not expect Llama 3 on a laptop to replace a frontier chat model at browsing, tools, or long documents. The context window on the file you downloaded may be shorter than the marketing for the family. If you paste a 50-page PDF into a model that holds a few thousand words, it will summarize the beginning and ignore the end, or it will refuse. Split the document.
Do not expect the app store of "agents" inside the runner to be private just because the chat is local. A tool that browses the web sends the query out. Read the tool before you enable it.
Do not fine-tune on the first weekend. Running the model is the skill. Changing its weights is a different project, described in the fine-tune guide only after you can delete a model on purpose.
Frequently asked questions
Q: Can a Mac run Llama 3 without an NVIDIA card?
A: Yes. Apple silicon uses unified memory, not a separate GPU card. The limit is that memory, plus the size of the file you download. Intel Macs can run smaller models and will feel slower. Try the small file before you plan around it.
Q: Which Llama 3 size should I download first?
A: The smallest instruct file the app offers in that family. Prove Wi-Fi-off answers on that file. Move up only when a real task fails because the model is too small, and only if the next file fits in memory with your other apps open.
Q: Is Ollama or LM Studio the right one?
A: Ollama if you like the terminal and a stable tag. LM Studio if you want a window and a model browser. Both can run the same family. One of them is enough.
Q: Does local mean nobody can see the prompt?
A: It means the prompt is not sent to a model host, if you have confirmed there is no cloud fallback. The text can still sit in app history, backups, and screen recordings. Local is a network fact, not a secrecy spell.
Q: Why is the answer worse than ChatGPT?
A: Because you are running a smaller model with no web search. Use it for drafts and private text. Use a hosted assistant, on purpose, when you need the larger one and the text is allowed to leave.
Q: Can I use it for coding?
A: Yes, for short functions and explanations, in a project you understand. It will invent APIs. The coding assistant guide covers tools built for a repository. A local chat is not that integration unless you wire it.
Q: The model downloaded but will not load. What now?
A: Read the error. It is usually memory, a broken download, or a tag that moved. Delete the partial file and pull it again. Do not install three more runners to see if one of them hides the error.
Q: How do I remove it?
A: Quit the app, delete the model in its own model list, and delete the app if you are done. Check the folder you wrote down in step 5. Empty the trash. If the folder is still huge, you deleted the app and not the weights.
You own the off switch
A local Llama 3 on a Mac is a file, a runner, and a test with the network off. Do those three and you have the setup. Skip the benchmark thread until a real task is slow or wrong. Then change one thing, the model size or the runner, and test again. The tools directory lists hosted options for the jobs this file should not pretend to do. The blog has the hardware notes if this Mac is simply too small.
