Apple Mac Mini M6 / M5 Pro: Local AI in a Compact Setup

00:00 Apple shipped four new chips in one week. There's a Mac mini with M6 and M5 Pro and a Mac Studio with M5 Max and M5 Ultra. Apple promises up to 13.5 times faster prompt processing and always-on agenda computing. It never published a single tokens per second figure anywhere, not one. That's the number that decides whether you buy. So, I ran the math Apple left out tier-by-tier.

00:23 Here's which box is worth the money and which one is a trap. Start with M6 because that's where the real news is. This is Apple's first chip built on a 2-nanometer process. The CPU runs 12 cores and for the first time Apple splits them three ways. You get two super cores, four performance cores, and six efficiency cores. The GPU has 12 cores now and each one carries its own

00:47 neuro accelerator built directly into the core. The neural engine jumps to a dual 16-core design. Apple rates it at up to double the peak neuro compute of M5. Bandwidth climbs to up to 170 GB/s, up from 153 on M5. Base memory is 16 gigs. The ceiling is 32. Hold on to that number because it's where this whole

01:10 video turns. That 170 figure comes with an asterisk. Apple splits that bandwidth number by configuration and the real footnote is simpler than the conflict in spec sheet rumor that circulated after launch. The base 16-gig M6 mini still moves 153 GB/s, exactly like M5. You only unlock the full 170 once you

01:34 configure 24 or 32 gigs of memory. So, the cheapest box doesn't get the bandwidth headline it sold on and almost nobody covering this mentioned it. I'm putting that on screen the same way we ran the correction card last time. Half the internet repeated 550 GB/s for a box that actually moved 273. Read the footnote before you spend $2,000.

01:59 The M5 Pro Mini is the interesting one. It goes up to 18 CPU cores and 20 GPU cores, and it takes up to 64 GB of unified memory. Data moves at 307 GB per second. You also get three Thunderbolt 5 ports instead of the M6 Thunderbolt 4. So, external storage stops being the bottleneck. It starts at $1,699,

02:23 and the base retail configuration is 24 gigs of memory with a 512 GB SSD. Both minis cost $100 more than the M4 machines they replace. The base mini has gone from 599 to 799 to 899, and none of that is because the chip got better. 32 GB sounds like plenty until you load a model. A 7 or 8 billion parameter model

02:47 at 4-bit quantization takes roughly 5 gigs, so it sits there comfortably. A 14B lands somewhere around 9 or 10. A 32B at 4-bit needs close to 20 GB before you've given it any context. Add a context window on top of that. Then add whatever macOS and your browser are already holding, and you're scraping the roof. That's the real picture of the

03:10 $899 box. It's a lovely little machine for small models and short prompts, and it is not an agent server. PC World put the M5 Pro tier just as bluntly. 64 GB will barely fit a 27B model you'd actually trust for work. That's the $1,700 tier, not the cheap one. AMD's Strix Halo refused to load GPT-OSS 120B under about

03:37 60 GB at all. So, a 64 gig M5 Pro can technically hold it and then has nothing left over for context. Meanwhile, the boxes it competes with start at 128 GB, which is where this comparison starts hurting. Memory size decides what loads, and bandwidth decides how it feels while it runs. At 170 GB per second, an 8B model at 4-bit should feel quick enough

04:02 that you never wait on it. Push a 32B onto the same machine, and you'll watch every sentence form. The M5 Pro's 307 GB per second roughly doubles that headroom, which is why the jump from 899 to 1,699 buys more than the spec sheet suggests. One thing kept nagging me while I built those memory tables. The whole reason

04:26 people want one of these boxes is to stop paying for models, and the honest answer is that no single box runs everything you need. I still keep Claude, GPT, and Gemini in rotation because each one wins at different job. A 32 gig mini isn't running any of them at full size. That's exactly the problem I built AI Master to solve, and it's my own platform. So, weigh the enthusiasm

04:49 accordingly. It's the multi-model workspace we actually run this channel on. This is the multi-model chat, and Claude, ChatGPT, Grok, and Gemini all sit in one window. I paste the same question about memory bandwidth into three of them side by side. Then I watch which one actually gets the arithmetic right instead of trusting a single answer. That's the same instinct driving this whole video. Don't take one number

05:13 on faith. Run it against something else, and you pay per token instead of stacking four separate monthly subscriptions to do it. The rest of the platform sits in the same window. Image, audio, and video generation all live in one place and come out cheaper than paying for each service separately. You can build consistent characters that look the same across every generation. Then you publish that character and let

05:37 it earn on the platform. Generated content gets shared through the feed, so it works for reach instead of sitting in the folder. Then there's Virable, which is the piece I'm most attached to. You and your team run a whole channel inside of it with AI agents writing scripts and descriptions, editing videos and images, and building thumbnails. The Academy sits underneath all of it with over 200

06:00 lessons, roughly 30 hours of material, and the community is past 13,000 people now. The case studies on the landing page come with screenshots and real creator testimonials. Those are pulled straight from their demo videos, not stock quotes written in an afternoon. There's an annual subscription with a discount and a 7-day money-back guarantee. So, if it doesn't fit your workflow, you take the money back.

06:23 Here's how you get access. Go to the link in the description and hit buy. Select the annual plan, fill in your details, and you'll get a confirmation email. Then log in and you're inside the same workspace I just showed you. The whole setup takes under 3 minutes. Link is below. Apple describes its AI performance almost entirely in relative multipliers. And this is the part every

06:46 other channel repeated without blinking. M6 is up to 13.5 times faster at LLM prompt processing in LM Studio than an M1 mini and 4.8 times faster than M4. M5 Pro gets up to 8.5 times versus M2 Pro and four times versus M4 Pro. Every single one of those is a ratio against

07:10 an old Mac. Multiply an unknown by 13.5 and you still have an unknown. Prompt processing is also the friendliest possible metric to lead with. That's the phase where the machine chews through your input before it says anything and it parallelizes beautifully on a neural engine. Generation speed, the number you stare at while text crawls onto the screen, is bound by memory bandwidth and bandwidth went up 11% from M5 to M6. So,

07:36 the headline number is the one that improved 13 times over 5 years and the number you actually feel improved by a tenth in a year. Apple picked the flattering half of the truth and shipped it as the headline. If you're running a business and juggling 10 different tools right now, what if I told you that one platform could replace all of them? This is Go Highlevel. You get access to pretty much all the features. Social

08:00 media planner, you can schedule posts, see your content calendar, and even repost content across multiple platforms with one click. Email marketing, Go Highlevel has a full email suite. You can create campaigns, build email templates with a drag-and-drop editor. Go Highlevel has a built-in calendar and booking system. This replaces tools like Calendly. You can create different

08:22 calendar types, one-on-one meetings, group calls, round-robin scheduling if you have a team. And here's the best part. Go Highlevel has a ton of pre-built workflow templates. I've linked a 30-day free trial in the description, not the standard 14 days, an exclusive extended trial. Here's where Apple's position gets uncomfortable. AMD's Strix Halo boxes and Nvidia's DGX Spark systems generally

08:47 start at 128 GB of unified memory. That's four times the M6 ceiling and double what the $1,699 M5 Pro can hold. AMD's Strix Halo box runs GPT-OSS 120B at around 34 tokens per second. The 128 GB Strix Halo box doesn't actually hand

09:11 you the full 128 GB for a model. The OS and overhead eat into it, so usable memory lands closer to 96 GB. A 64 GB M5 Pro still can't load a 120B model in any usable way, even after that correction. It never gets the chance to run it slowly. It just doesn't run. The PC side deserves a caveat before you go spend

09:35 the money you just saved. A Strix Halo box hands you the memory and then it asks for patience in return. The software stack around AMD's accelerators is still thinner than CUDA and a lot of the tooling quietly assumes you're on Nvidia. Nvidia's spark boxes solve that problem and charge you for solving it. Apple's pitch is that everything just runs quietly on the machine you already understand. That's worth real money. Mac

10:00 Studio is Apple's serious answer here and it costs accordingly. The 512 gig Ultra runs models nothing else on a desk can hold at $5,499 before you've touched the memory slider. So, the lineup makes sense at the top and stops making sense at the bottom. The Mini is the machine most people in this audience will actually buy and it's

10:23 the one with the ceiling problem. Stack up a typical creator's monthly bill and the subscription math gets real fast. You'll find something like $20 for a chat plan and another 20 for a coding assistant. Image generation adds even more on top. Then API charges stack on top the moment you build anything real. Call it 60 to 200 a month depending on how deep you're in. Against that, a

10:46 $1,699 M5 Pro pays itself off in about nine months at the high end. At the low end, it takes well past two years. Electricity barely registers here. Running that box hard for a few hours a day costs single digit dollars a month in most places. The arithmetic checks out. What you actually get on the other side is the real question. Your $1,699 box runs a good 27B model and a good 27B

11:14 model is not GPT-5 or Claude. It's genuinely useful for summarizing, drafting, tagging, and classifying. It handles an agent loop overnight and anything you'd rather not send to a third party. You still reach for the frontier models when the work has to be right the first time. So, the subscriptions don't die. The volume work moves local and gets free while the hard work stays on the same paid models as

11:39 before. Your bill shrinks. It doesn't disappear. Anyone selling you the clean replacement story is skipping that step. There's one workload where a local box wins outright. Bulk jobs that run overnight are the whole reason to own one. You transcribe and tag a year of footage or rewrite a thousand product descriptions. Maybe you leave an agent looping through your own files while you

12:01 sleep. Metered API pricing punishes exactly that shape of work because you pay for every token whether the output was useful or not. Local inference costs you electricity. So, the real comparison here is hardware against your API bill. And that bill grows every time you get ambitious. The price ladder is the other half of the story and it isn't really Apple's doing. Memory prices have been

12:24 climbing through a genuine shortage with analysts tracking server DRAM up in the mid teens percent in a single quarter. Those increases are forecast to run into 2027. Tim Cook called it a hundred-year flood on his final earnings call. That's how a mini went from 599 to 799 to 899 in about two years while gaining a hundred dollars again this round. You're

12:47 not just buying a chip. You're buying memory at the worst moment to buy memory. This uses the same rating system as the GPT 6 episode. Confirmed means it shipped or Apple said it directly. Credible is a reporter with a real track record. Then there's possible where the supply chain is talking and speculative where somebody just wants clicks. The credible tier here is the reason you

13:09 might not want to buy anything today. Mark Gurman reported in June that Apple is skipping M6 Pro, M6 Max, and M6 Ultra entirely and jumping straight to the M7 generation. Apple has never skipped a whole high-end tier before. That's why this year's studio carries M5 Max and M5 Ultra instead of M6 anything. And now the lineup makes sense. The M7 timeline

13:35 is the part that matters for your wallet. Same reporting puts base M7 in the first half of 2027 at around 240 GB per second with M7 Pro and Max landing late that year. M7 Ultra is penciled in for 2028, designed for up to 1.5 TB of unified memory. Internally, Apple frames it as approaching Nvidia's Blackwell

13:59 class. Write that one credible on timing and speculative on the performance framing because a target 18 months out is a wish with a roadmap attached. Below that sits the genuinely speculative pile. There's talk of an M8 on a 1.4 nanometer process around 2028, Apple's own AI server silicon, and a Mac Pro that comes back from the dead. That's

14:21 fun to watch, but none of it is something you plan a purchase around. There's a software story underneath all this hardware, and it's not a happy one. The co-creator of MLX, Apple's own machine learning framework, left for Anthropic in February, [music] and he wasn't alone. Around a dozen AI researchers have gone, including the person running foundation models. At the same time, Apple reportedly pays Google roughly a billion dollars a year for a

14:45 custom Gemini to power Siri. So, the company selling you on-device intelligence is licensing someone else's cloud model to run its own assistant. That's not fair for the hardware. It does tell you where the confidence is. Practically, that shows up in one place. Fine-tuning on Apple's Metal back-end is flaky enough that experienced ML people quietly avoid relying on it. Treat these

15:08 machines as inference boxes. They run models beautifully, and they are not where you train or tune one. If your workflow needs tuning, you're buying Nvidia, and no amount of unified memory changes that this year. Start with the one I'd skip. The M6 Mini is a lovely little computer and the wrong purchase for local AI. 32 GB is a wall you hit in week one and no software update moves

15:32 it. There's a good reason to buy it. You want a cheap, silent Mac that also happens to run an 8B model on the side. You'll be happy with it for years. There's a bad reason, too. A keynote slide said 13.5 times faster prompt processing. That's not why you buy a computer and you'll be shopping again by spring. The chip is not the limitation here. The memory ladder is. The M5 Pro

15:56 at 64 GB is the conditional buy and the condition matters. It makes sense if you want an always-on box that isn't your daily driver and if you already live inside Mac OS, it makes even more sense when the models you actually run top out around 27 to 32 billion parameters. That covers a lot of real work. It stops covering it the moment you want the big open weight models. At that point, 128

16:20 GB of PC memory costs less than 64 GB of Apple memory. The Studio is the one purchase in this lineup I'd call straightforwardly good with a caveat about who it's for. It's a workstation price for a workstation and most people watching this don't need one. If local models are how you earn though, the M5 Max is where I'd start looking because mini money was never going to get you

16:44 there. You skip the ceiling problem entirely and the machine stays useful long after M7 exists. The Ultra above it is a different conversation and a different budget. Anyone who needs half a terabyte of unified memory already knew that before this video started. Then there's waiting and waiting is where most people will land. If Gurman is right, the next genuinely new

17:07 high-end Apple silicon is M7 in 2027 and this Studio generation is a placeholder wearing last year's chip. So the patient move looks obvious. The catch is that memory prices are forecast to keep climbing into 2027. So, a year of waiting probably buys you a better chip at a worse price. Wait if you're curious. Buy now if the box has a job on Monday. So, here's where I land. The M6

17:32 Mini is the most overrated AI machine Apple has ever marketed. The missing tokens per second number bothers me more than the memory ceiling does. Apple shipped four chips in one week without printing that figure once. This box hands its GPU 96 GB of memory. Your laptop hands its GPU about eight. That single gap is the whole reason you rent intelligence for $20 a month instead of owning it. The HP Z2 Mini G1A holds 128