Don't Buy a Mac for AI Until You Answer These 3 Questions

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Don’t Buy a Mac for AI Until You Answer These 3 Questions · Channel: Hyperautomation Labs · 09:58 · Watch on YouTube

Topics: AI agents, Model inference and optimization, AI infrastructure and hardware, Local and open-source models, Open source and software ecosystems

Summary

The video introduces a free Mac picker that recommends a machine based on what you want to do, the speed you need, and the Mac you already own. It compares local AI workloads and measured model speeds with Apple Mac prices, memory, and bandwidth, emphasizing that an existing Mac may already be sufficient. Three examples show when to keep a current machine, when an upgrade is only a speed improvement, and what to consider for running agents or larger open models. The central advice is to check whether your current Mac can run your work at the speed you need before buying another one.

Key points

The picker starts with three questions: what you want the Mac to do, how fast it needs to be, and what machine you already own.
Workloads are grouped into four broad classes, from chat and private documents to agents and the largest open models.
The page ranks 37 open models for the selected workload without requiring users to choose a model by name.
The video defines reading pace as 8 tokens per second or more, comfortable as 15 or more, and fast as 30 or more.
The presenter measured seven models on a 24 GB M4 Pro and says the page’s formula correctly reflected their speed bands, including a model that spilled beyond graphics memory.
For private documents and drafts at reading pace, the example MacBook Air M4 with 16 GB is a keep, with Gemma 4 12B estimated at about 10 tokens per second.
For coding at reading pace, the presenter’s M4 Pro can stay; a Mac Mini M5 Pro with 48 GB is presented as a speed upgrade to comfortable pace, costing $2,299.
The video argues that memory bandwidth, rather than the chip’s release year, determines writing speed, so some used Macs can outperform newer models.
The agent-running example recommends a 128 GB M5 Max Studio, while also showing a used M1 Ultra Studio as a cheaper option; some larger-model speeds are explicitly formula-based, not measured.
The picker lists Mac and memory combinations, model sizes, estimated or verified prices, speed estimates, and labels such as runs tight, too slow, or no fit.

Chapters

Start	Chapter	Summary
00:00	A three-question Mac picker	The presenter introduces a free tool that bases Mac recommendations on workload, required speed, and the computer the viewer already owns.
01:06	Workloads, speed targets, and measured results	The picker maps everyday tasks to model classes, defines three speed bands, and compares its estimates with model runs measured on an M4 Pro.
03:26	Three examples: keep, upgrade, or buy	The presenter applies the picker to an Air owner, his own M4 Pro, and a prospective agent user, showing the recommendations and costs.
06:10	Why bandwidth and used Macs matter	The video compares memory bandwidth across Mac chips and explains how the picker evaluates used machines and larger-model configurations.
07:51	What the picker shows and the final rule	The presenter describes the tool’s detailed tables and concludes that viewers should keep a Mac if it meets their workload and speed needs.

Quotes

What do you want the machine to do? How fast does it need to be? And what do you already own? — 00:00

That is the answer most people who click on a Mac video actually need. — 04:29

Writing speed is set by memory bandwidth, not by the year on the box. — 05:38

If the Mac you own runs the model for your job at the speed you asked for, keep it, install it tonight and buy later for speed if you still want it. — 08:21

Concepts and entities

Apple, MacBook Air M4, MacBook Pro M4 Pro, Mac Studio M5 Max, Qwen3.5, Gemma 4, DeepSeek V4 Flash, GLM-5.3-Flash, largura de banda da memória, Claude Code

Full transcript (original language)

00:00 Apple ships the new Macs on the 22nd and I am going to stop you from buying the wrong one. Every guide starts with model buckets and memory tiers. I built a page that starts with three questions instead. What do you want the machine to do? How fast does it need to be? And what do you already own? The first thing it told me was to keep the Mac I already have. By the end of this video, you will have one Mac with Apple’s real price, one model command to

00:32 run on it, and the month your AI subscription would have paid for it. Every speed on that page was checked against runs on my own Mac, and the receipts are on screen. The page is free, no sign up, and it is one comment away. Here is why I did not start from the models. Watch the Apple page. Max Studio 2499 Click the bigger chip 399 64 GB 3,499 128 599

01:06 and the Ultra 5,499 to start and 10,799 with the top chip and 256 GB. Those are the numbers you’re actually deciding about. And the guide’s answer with model names you have never typed. So the page asks what you would type instead. Question one, what do you want it to do? Four answers. Chat, summaries, drafts, and private documents, which is small models and the fast small expert models. contracts, spreadsheets, and code fixes,

01:42 one task at a time, which is the 27 billion class, the class I tested against a $200 coding plan on real work. Agents that run for a while across many files, which is the 100 billion class, and the biggest open models, cloud class at home. Behind those four answers sit 37 open models with published sizes. Gwen, Gemma, GPTOSS from OpenAI, Meta’s new 30B, Mistral Granite from IBM, Nvidia’s Neatron, Coher, Deepseek, GLM, Kimmy. You never have to name one. The

02:15 page ranks them inside your answer and picks the strongest one that fits. Question two, how fast does it need to be? Reading pace is eight tokens a second and up. Fine for chat and short answers. Comfortable is 15 and up, which is coding and long documents without waiting on it. Fast is 30 and up for agent loops that make call after call. And here is where the receipts come in. I measured seven models on my 24 GB M4 Pro. Quen 3.5, 9B, 33 tokens a second,

02:49 Gemma 4 12B 26, the 27B at 3 bits 12.4, 4 and the 17 GB default build of the same model 0.3 tokens a second because it spilled out of graphics memory. The pages formula landed in the right speedb band on all seven including that spill. That is the whole reason I trusted above 24 GB where I could not measure. Question three, your budget, your subscription and the Mac you already own. Every Apple silicon Mac is in the list, 47 of them, from the 2020 M1 Air

03:26 to the M5 Ultra Studio, each with the memory it shipped in and Apple’s own bandwidth number for its chip. The prices for the new machines came straight out of the configurator today. Mac Mini M6, 899, 1,99, 1,299 for 16, 24, and 32 GB. Mac Mini M 5 Pro 1,699 2,299

03:56 2,699 for 24 48 and 64. Now let me run three people through it. Person one owns a MacBook Air M4 16 GB and wants private documents and drafts at reading pace. Chat and documents reading pace MacBook Air M4 16. Verdict: Keep it. Gemma 4 12B about 10 tokens a second and it fits with 2 GB kept back for the apps you already run. One command, install it

04:29 tonight, $0. That is the answer most people who click on a Mac video actually need. And it is the one the buying guides never give. If this page is already worth more than a click, hit subscribe. It’s free. And if you want to back the work, hit join here or the supporter button on Facebook. Now, my own Mac person two is me, MacBook Pro, M4 Pro, 24 GB, contracts, and code fixes, one task at a time. At reading pace, the

05:03 verdict is keep it. Quen 3.8, 8 27B the 13 GB build about 13 tokens a second on the page and I measured 12.4 four on the desk. That is the model that passed 14 of 15 real coding tasks against my $200 plan in August. Now watch what happens when I ask for comfortable. Buy this one. Mac Mini M5 Pro 48 GB 2,299 running the same model at about 18 tokens a second. And the page says the

05:38 honest thing under it. My Mac already runs this at 13. So, this is a speed upgrade, not a can or cannot. $2,299 divided by a $200 plan is 11 and 1/2 months. That is the decision in one card. Here is the part that changes which Mac you buy. Writing speed is set by memory bandwidth, not by the year on the box. The chip reads the whole model out of memory for every single token. So, the

06:10 gigabytes per second is the speed limit. The new M5 Pro moves 307 GB a second. A 2021 M1 max moves 400. The new M5 max moves 460 on the base chip and 614 on the 40 core. A 2022 M1 Ultra moves 800. So, a used studio from 3 years ago can write faster than a brand new mini. And the page shows you which used Macs pass your answers so you can price them on the used market that day. The one number

06:43 nobody has measured yet is the new ultra 1.2 tab a second on Apple’s page. Until someone measures it, the page discounts it and labels it. Person 3 has no Mac yet and wants agents. Many files, long context, comfortable speed, no budget limit. Buy this one. Max Studio M5 Max with the 40 core chip 128 GB 5,99. It runs Quen 3.5 122B about 60 tokens a second because

07:16 that model only wakes 10 billion parameters per token. Under it, the used row, an M1 Ultra Studio with 128 GB does the same job at about 50. Now, the biggest open models. The page moves to the ultra with 256 GB 9,499 running DeepSeek 54 flash. And if you’re willing to raise the Mac OS graphics memory limit, which Apple does not support, the same machine holds GLM 5.3 flash at 4 bits above 24 GB. These speeds are formula, not measurement. The

07:51 page says so on every row. If you would rather see everything, scroll down. Every Mac against your answers. 120 machine and memory combinations, each with the model it would run, the speed, and a stamp that says runs tight, too slow, or no fit. Filter to the ones Apple sells today or to the ones that pass. Below that, every model with its download size, the bytes the chip moves

08:21 per token and where the size came from. The Olma library or on slots files dated and the receipts table seven measured rows against the formula. Prices with a tilda are estimated from Apple’s tier pricing and marked that way. Nothing on the page pretends to be more certain than it is. So, here is the rule and it is the whole video in one sentence. If the Mac you own runs the model for your job at the speed you asked for, keep it, install it tonight and buy later for

08:54 speed if you still want it. The Air keeps. My M4 Pro keeps at reading pace and upgrades for 2,299 if I want comfortable. The Agent Runner buys the 128 GB Studio or a used ultra for less. Three people, three questions, three answers, and not one of them started from a model name. Real talk, 90% of you watching haven’t subscribed. On Facebook, subscribe is the supporter button. That is what keeps the lights on here. On YouTube, it’s

09:27 free and it tells the algorithm to send you more. Subscribe to support us. The picker is linked in the description and the card that goes with it. Three questions. The rule, the verified prices, and the bandwidth ladder on one page is free too. Comment the word which MacA written as one word whi mac and bot sends it to you. If you’re just getting started, my beginner guides are linked below. The complete Claude code guide, the OpenAI codeex guide, the Claude

09:58 Cowwork sales guide, and the Claude architect prep kit. Follow along on Instagram and Facebook for byte-size AI and the deep dives live here on YouTube. Answer the three questions before you answer apples. See you in the next one.