This Mac Mini runs all of my AI agents. It sits right here on my desk and it’s on all the time. I paid $700 for it last year. But the Mac Mini is only the body. Most of the AI models that do the actual thinking, I rent every single month from Anthropic and Open AAI. And those plans go from $20 a month to $200. OpenAI even has a new $500 subscription. So, I wanted to know, can a Mac Mini replace a $200 AI subscription? So, I actually tested it. And in this video, you’ll see what a Mac Mini can actually run, where it performs better than any subscription, at what price point it starts to make sense, and where it falls apart. Let’s get started. Before we answer the question, can a Mac Mini replace a $200 AI subscription? I think it’s important we understand what is it actually that we’re paying for. An AI agent is actually two parts. There’s the body. That’s the computer it lives on. It’s always on and it runs all our jobs. And then there’s the brain. That’s the AI model it sends everything to. So it can think. And right now my body is this Mac Mini and my brain is clawed and GPT running in someone else’s data center and I pay for every bit of thinking it does. And here’s something nobody mentions. When you pay $200 a month, you’re not paying for one thing. You’re paying for three different things. Number one is the model itself, the actual brain. On the cloth side, that’s Opus and Fable. on OpenAI. That’s GPT Astra for example. Number two, you’re paying for somebody else’s computer to do the reading because every time your agent asks the model anything, a whole data center has to read and process everything first. And number three is the allowance. In other words, how much you’re allowed to use before it cuts you off. But here’s the twist. A Mac Mini can technically do all three of those things. It can be the model. It can do the reading itself. And there’s no cap. So on paper, this little box replaces the whole subscription. However, can that one small box do a job that normally requires a data center full of GPUs? And to answer that question, we need to look at the most important spec on any machine. And it’s not the chip, it’s memory. To run a AI model on your computer, the whole model has to fit in your memory. In other words, your RAM. If it doesn’t fit, it doesn’t run. It’s as simple as that. And on a Mac, that memory is shared. The same 16 gigs runs MacOSS, your browser, your apps, and the AI model. If I open my activity monitor and click on memory, you can see MacOSS and my apps are already almost using 14 gigs of the 16 gigs we have. And I haven’t even loaded the single model yet. On top of that, MacOSS only lets the part of the chip that runs AI use a chunk of that memory. On 16 gigs, that’s somewhere between 10 to 12 gigs for the model. Now a model’s size is measured in parameters which is basically how many connections its brain has. More parameters usually means smarter but it also means much bigger. And those big 27 billion parameter models or you see 27B at the end of their name that you keep seeing in every YouTube video, they’re 16 to 18 gigs. So they don’t even fit. But what if I told you you can predict how fast a model will run before you even download it? Every time the model writes one token, which is about three quarters of a word, the chip has to read the whole model from memory every single time. So the speed is basically how fast your chip reads memory divided by how big the model is. Now, Apple says that my M4 Mac Mini reads memory at 120 GB a second. So, a model that’s around 7 gigs should land somewhere around 15 to 20 tokens a second. But let’s see if that’s actually even true. I already showed you how to install Olama in a previous video. So, if you want to watch it after this video, I’ll link it up here and also in this video’s description. Olama is basically the free app that runs AI models on your own computer. But what I will do is open a terminal page and type in Olama Run Quinn 3.59B verbose. The verbose part just makes it show you the speed at the end of every prompt and I will ask it to summarize the Mac Mini Wikipedia page for me. Then I will actually paste the same exact prompt to Claude as well so we can test them both side by side. Oh, and this reminds me, you can type in Olama PS and if it shows you 100% GPU just like it shows for us, then the model fit. So, we’re good to go. But it if it shows you CPU anywhere, then the model is splitting out of your memory and it’s going to crawl. But that being said, let’s run both of the tests and see how both perform. There you go. Claude has already started answering as soon as we put that prompt in. That wasn’t even 3 seconds of wait time.
while our Mac Mini took 2 minutes and 20 seconds to go through this prompt. And as expected, it went through 13 tokens per second. But that test was only on a small prompt, and my agent doesn’t just chat. And that’s where things start to get interesting. Almost every speed number you see online is measured on a short question like write me a poem. But that’s not how an agent talks to its brain. Every time my agent asks the model anything, it sends a lot more than just my message. It sends it its instructions, its memory, the list of every tool it can use and the whole conversation so far. And before the model can write a single word back, it has to read all of that first. Some people call that prefill, which is just a fancy word for the model reading everything before it answers. Now, fair warning before you try this yourself. On a Mac with less than 24 gigs of memory, Olama only gives the model about 4,000 tokens of memory by default. And that’s very tiny. An agent needs at least 64,000. So, what you do is you go into O Lama settings and drag that context bar. Otherwise, your agent forgets what it’s doing halfway through the job. So, to test this next question out, let’s give it a much longer prompt about the size my agent actually sends and start the timer. There you go. Claude has already answered. And as you can see, this prompt took a whole lot longer to process. And that right here is the trade with local models. The model itself is free to run, but you pay in waiting time instead. And you know what they say, time is money. So now the real question is, is the wait even a problem if the job runs while you’re not watching? So we know it runs and we know the agent has to wait longer on the big jobs. But is that weight actually a deal breakaker? Here’s the thing. For a job that runs on a schedule while I’m asleep and away from my desk, the weight doesn’t cost me anything. A morning brief can take 10 minutes instead of two. I’m not sitting there watching it and waiting for it. So, the weight isn’t really where it falls apart. The real limit is quality. And honestly, so far, we’ve only been scratching the surface because we’ve only looked at the small models that fit inside my little box. So, let’s zoom all the way out and look at every model side by side. This is the artificial analysis chart. It takes basically every AI model out there and scores how smart it is in one single number. And right at the top of this chart, you’ve got the best paid models, the clawed and open AI models scoring around 58. Now, here’s the part that surprised me because the best open models, and by open I mean the ones that are completely free to download, aren’t actually that far behind. They’re sitting at around 44 to 46. So, the free brains really are closing in on the paid ones. But, can you actually run one? Well, the best open model on that entire chart is over a trillion parameters big. Remember, a parameter is just one connection in the model’s brain. And a trillion of them is massive. And remember the rule from earlier, the whole thing has to fit in your memory. Even after you crush that model down to a fraction of its size, it still needs around 236 gigs of memory just to load. The biggest Mac Mini that money can buy is 64 gigs. And that costs about $2,700 or $3,500 if you’re in Europe like me. and we’d still need something nearly four times that size just to load it. So to run the good open brain at home, a Mac Mini isn’t even an option. You’d need a Mac Studio with $256 gigs of memory. That is about $9,500. So let’s actually do the math on that. $9,500 against a $200 a month subscription. That machine takes about four years just to pay for itself. And even after those four years, it’s slower than the subscription. And because you had to squeeze the model down to make it fit, it isn’t even as smart as that chart says it is. So chasing a frontier brain at home cost you 10,000 grand in 4 years, and you still lose on both speed and smarts. Which brings us back to our actual Mac Mini and the small model that does fit on it. Because odds are when you run that little model locally, you’re not replacing $200 of clawed opus. Nor should you be expecting to do that. You’re replacing what that exact same model costs to just rent online. And that’s almost nothing. About 10 cents for a million tokens to be exact. So even if your agent runs all day, you’re looking at a few dollars a month, not $200. That small brain lives way further down the same chart. So on the hard stuff, it’s going to lose, but on a clean, simple job, it’s absolutely fine. So the money was never the reason to do this. The real reason is something a subscription can’t give you at any price. See, cheaper isn’t even the real story here because there are three things this little box does that no subscription can at any price. The first one is privacy. When the brain runs on this Mac, your data can never leave this room. Your inbox, your calendar, your documents, none of it gets sent to a company. And that was the number one reason people say they want to run a local model in the first place. The second is that there’s no cap. There’s no 5h hour limit. There’s no weekly limit. There’s no message telling you to come back later. It just runs all day and every day as much as you want. And the third is the running cost, which is basically nothing. Apple says this Mac Mini uses 4 watts sitting idle and 65 watts flat out. Here in the Netherlands, electricity is around 26 cents per kilowatt hour. So sitting idle all month, that’s about.74. And even if it ran flat out 24 hours a day, the whole month, that’s about €12 a month. So here’s where we land. On the hard stuff, the paid brain wins. And it’s not even close. But on the private jobs, the repetitive jobs, and the things that run while you sleep, the free one isn’t just cheaper, it’s actually better. Okay, so say you want to do this. What do you actually buy? The new base Mac Mini, the M6, is $900. It’s faster than mine, but it still has 16 gigs of memory. So, it runs the exact same size models I just tested. Nothing bigger fits. The one that actually changes things is the M6 with 32 gigs. That’s $1,300. And that’s where the bigger, smarter models finally fit. the ones that don’t even load on mine. And the jump from 16 to 32 gigs is $400, which is two month of a $200 plan. And then there’s the M5 Pro chip with 64 gigs from $2,700. It reads memory about twice as fast, so everything runs quicker. But that’s for people making local AI their actual hobby. So, if you’re buying one just for this, 32 gigs is the floor. So, back to the question. Can a Mac Mini replace a $200 AI subscription? No, not on its own. And honestly, it was never really trying to. But here’s the smart money move, and nobody selling you a Mac Mini will tell you this one. Before you buy anything, just spend one week reading your own usage. If you’re not actually hitting your limit, drop to the 20 or the $100 plan today. You’ll have access to the same exact models you already pay for, but you’ll have more money back in your pocket with zero hardware or zero new things to change. Because what this Mac Mini actually is isn’t a replacement. It’s the best home for always on agents that you own like Hermes agent or OpenClaw. The brain that lives here handles the private stuff and the jobs that run on a schedule. And when a job is too hard for it, it hands it off for the paid brain for that one task. So tell me in the comments, would you trust a free local brain with your inbox or are you keeping the subscription? And I put together a free sheet with which model to run on which Mac, plus the exact settings I used. It’s the first link in the description. And if you enjoyed this video, then leave a like. Or if you’re new to my channel, then subscribe because I’ve got a ton more content like this coming your way. Oh, and would you look at that? The algorithm gods seem to think that you’re really going to enjoy this next video as well. So, click it and I’ll see you there.