Mac Mini 16GB vs 32GB for AI Agents: OpenClaw Uses 1GB

Vídeo em Inglês, publicado pelo canal Mind The Prompt.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video discusses how to choose the right Mac mini for running AI agents by analyzing whether the AI processing happens in the cloud or locally on the device, examining memory requirements, token processing speeds, and cost-effectiveness.

Key points

• Buying guides often recommend a 32GB Mac mini for AI agents, but hardware needs depend heavily on whether the AI model runs locally or via cloud services.
• If an AI agent routes its thinking to the cloud (such as using a ChatGPT login), a base model Mac mini with less memory is sufficient since the hardware only acts as a host.
• Running dense local models requires significant memory; for instance, a 27 billion-parameter Qwen model needs a 32GB machine to fit into memory.
• Agent tools send large amounts of initial tokens (instructions to themselves) before answering, which can cause noticeable delays and slow response times on local hardware.
• Mixture of experts models, like Google's Gemma, offer much faster token processing speeds compared to dense models on the same 32GB hardware.
• Determining the hardware requirements starts with checking the software's manual and identifying whether the model processing is rented or local.

Topics

Mac miniAI agentsHardware specificationsLocal AI modelsCloud thinkingMemory and tokensCost analysis

Video transcript

Transcript from available captions; it may contain errors.

I was in the vet's waiting room, with the dog, sat firmly on both my feet, when a friend messaged me. This friend pays for ChatGPT and for Claude, has never once run a model on their own computer, and had decided they wanted one of these AI agents, the kind that lives on a small computer at home and gets on with things while you're out. The message said, which Mac mini? I replied in about 4 seconds, 32 gigs, $1,300. That's the answer in nearly every buying guide going. Skip the $899 one, 16 gigs is a trap, get 32. And I believed it enough to have that exact machine up on my own phone, one double click of the side button from being mine. Then they called the dog in and the phone went in my pocket. That evening, I did what I should have done first, I read the agent's manual, and 32 gigs turned out to be the wrong answer for my friend, and for me, for a completely different reason, it was only half an answer. The guides have good reasons, and they deserve a hearing. If you want the model itself running on the Mac, memory decides everything. A model is a file, and the whole file has to sit in memory. On a Mac this size, by the usual rule of thumb, the system keeps about a third back, so 16 gigs is really under 11. The model people recommend most for agent work is a 27 billion-parameter Qwen. Its file is 16.5 gigs. On the base machine, it's nearly 6 gigs too big before you've typed a word. On 32, it fits. Apple charges $25 for each extra gig, so that's $400 more. And the regret is genuine. I went through 34 owners describing their own Mac minis, 9, wished they'd bought more memory. All 9 were running a model at home, or trying to. So if you bought 16 and it's struggling, you've read the situation correctly, but the manual I'd opened wasn't about models at all. The agent, my friend wanted, is OpenClaw, nearly 400,000 stars on GitHub, and the tool that set off this year's rush on Mac minis. Its documentation has a hardware line, absolute minimum, 1GB of memory, recommended 2 1. Not 16. There's even a question in the FAQ, do I have to buy a Mac mini to install this? And the first word of the reply is no. For most people, it recommends a small rented server. I'd spent the afternoon looking at the explanation without seeing it. At the vet, the person on reception finds your file, books you in and takes your money. Busy all day. But they don't diagnose the dog. That happens in another room. An agent is reception. It reads your messages, opens files, clicks buttons. The thinking is done by a model, and on a fresh install, OpenClaw sends the thinking out to a cloud model, through a ChatGPT login. So which Mac mini for AI is two questions in one coat. And this one comes first, where does the thinking happen? Take my friend's case, the thinking is rented. Then the Mac mini is only the box that stays switched on. The base model has 16 times the memory the agent asks for. The $400 upgrade by 16 more gigabytes that nothing will ever touch. And it is a lovely box for the job. Apple says it idles at four watts. That's 53 cents of electricity a month. There are honest reasons to want a Mac as well. Your messages, your Apple account, a home internet address. But look at what it's competing with. One guide prices the sort of small server those docs recommend at about $6 a month. The base Mac mini is 12 and a half years of that server. The memory upgrade alone is another five and a half. Someone even ran two always on coding agents on a Raspberry Pi and reported them using about 900 megabytes. So for the rented half, the right Mac mini is the cheapest one. Or the computer you already own. Then there's me. I want the thinking at home, so surely 32 gigs is my machine. Enter a developer called Eishan Lawrence, who did something I love. He took 12 agent tools, gave everyone the same model and the same 12 small coding jobs, and counted the tokens. Above all the tokens each tool sends before you've typed anything. It's instructions to itself. The leanest tool sends 700. Claude Code, about 4,500. OpenClaw, about 26,000. It's one run per pairing on small jobs, so it's a sketch. But owners report the same shape. 30,000 to 50,000 tokens a message. Now put that on the 32 gig Mac mini. Community results have that Qwen model reading about 313 tokens a second. 26,000 tokens from cold is 83 seconds. Only then does it start to answer. Well over a minute of silence to say hello. The writing is no quicker. An owner called Hans-Dieter Zucht ran 16 models through his own 32 gig M6. 3 runs per prompt and published every result. That Qwen model writes at about 11 tokens a second. He also saw the machine reach 95 degrees doing it and wrote his own fan controller. I admire that enormously. 11 can't be tuned away. Divide the memory speed Apple quotes 170 gigabytes a second by the file 16.5 gigs and the ceiling comes out at about 10. The measurements sit right on it. That machine is already as fast as its memory allows. Then there's room. Another agent Hermes says in its own guide that it needs 64,000 tokens of conversation as a minimum. For this model that's about 4.5 gigs on top of the file. 21 gigs altogether against a share of 21.3. A quarter of a gig to spare. On the upgraded one. What cheered me up is that the same machine has a far better answer and it costs nothing. There's a kind of model called a mixture of experts. It's big on disk but only a small slice of it works on any one word. Google's Gemma is one, 26 billion parameters, about 4 billion awake at a time. On that 32 gig Mac mini it reads over 1,300 tokens a second. The 26,000 token opening takes 19 seconds, not 83. And it writes at about 48 tokens a second. Against 11. Same box. Four times faster. In both directions. It has a price. Some owners find these models follow instructions less carefully than the dense one. But it's the gap between an agent you use and an agent you wait for. And one footnote while we're on speed. Apple's headline nearly five times faster prompt processing than the previous Mac mini was measured on the 32 gig model. The 899 machine isn't in the footnote. So what does thinking at home save? Begin with what the cloud charges. A benchmark of eight agent tools on a cheap hosted model put one finished task at between 3 cents and 20 cents. At 20 cents, the $1,300 Mac has paid for itself after about 6,700 tasks. At 3 cents, 46,000. Suppose your agent finishes 20 jobs a day. That's my guess, not a measurement. Then it's somewhere between 11 months and six and a half years. And for my friend, the upgrade replaces no tokens whatsoever. $400 saving nothing breaks even never. The real money risk is somewhere else. One OpenClaw owner found his agent burning 50 million tokens a day. A large part of it was a heartbeat. Every 30 minutes the agent woke up, reread a 225,000 token history and answered, heartbeat OK. Nearly 11 million tokens a day to say I'm fine. So you don't run a model at home to get rich. You do it so your files never leave the house. That's a good reason. It just isn't a financial one. Now, I've drawn that line very tidily and real buyers are messier. Of those 34 owners, 20 wanted models at home and only four were cloud only. That's a crowd that posts about memory. So it leans one way and nobody has counted the rest. I can't tell you what most buyers do. Nobody can. The 83 seconds is a cold start, worked out from a short test. With the prompt cached, it's far quicker. The third the system keeps back is a rule of thumb. A setting raises it and nobody has confirmed it on this machine. And I couldn't find one measured speed on the 16 gig model. Not one. So here's a check for tonight. Open your agent's settings and find the line that names the model. If it names a company, the thinking is rented and memory is beside the point. If it says Ollama or LM Studio, the thinking is yours and memory is the point. Then in OpenClaw, type slash context list and see how big your opening really is. My friend did get a proper answer in the end. Four sentences. Decide where the thinking happens before you look at a single price. If it's rented buy the base model or nothing and set the heartbeat to zero until you need it. If it's at home, get 32 gigs, run a mixture of experts model and trust only two thirds of the memory you paid for. And for the big dense model with no waiting, the M6 Mac mini is the wrong Mac. That's 48 gigs on the Pro chip, $1,000 more with nearly twice the memory speed. Which Mac mini for AI was never one question. Size the box for the thinker, not for the messenger. 32 gigs is good advice. It was written for somebody running the model at home, then handed to everybody. What I'd change is the order I read things in. The software's manual first, the hardware's spec sheet second. My friend bought nothing. Their agent runs on a laptop from a cupboard. And that afternoon at the vet, the only one in our house who cost $400 that week was the dog.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal. Publicado no YouTube em 2026-10-17.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.

About OpenClaw and its requirements

OpenClaw is an open-source personal AI assistant created by Peter Steinberger, written in TypeScript and released under the MIT license. It runs on the user's own machine through a Gateway and answers on channels such as WhatsApp, Telegram, Slack, Discord and iMessage, using either hosted or local language models.

According to OpenClaw's installation guide, it needs Node 24.16+ or 26.1+ (Node 26 is recommended) and runs on macOS, Linux and Windows. For local models, the documentation says memory needs depend on the model weights, context size, runtime and other work on the host. The managed llama.cpp setup checks available RAM, supported GPU memory and disk space before downloading; its curated recipes use a 64K context, and the smallest has an 8 GiB host-memory floor. Other supported local backends include LM Studio, Ollama, MLX, vLLM and SGLang.

Fact sheet

• Software: OpenClaw
• Type: Open-source personal AI assistant
• Creator: Peter Steinberger
• License: MIT
• Runtime: Node 24.16+ or 26.1+ (Node 26 recommended)
• Operating systems: macOS, Linux, Windows
• Models: Hosted or local
• Local model memory: From 8 GiB (smallest managed llama.cpp recipe); larger models need more
• Documentation: docs.openclaw.ai/install

Read also

• Build Your First AI Assistant on a VPS in 15 Minutes With OpenClaw
• How to Use OpenClaw as Your Personal AI Coworker (Beginner Guide)
• Meta unveils Muse personal AI assistant

---

Sources and references: OpenClaw Docs – Install · OpenClaw Docs – Local models · OpenClaw repository on GitHub · Wikidata – OpenClaw (Q137916205) · video by the Mind The Prompt channel on YouTube.