Source: raw/Can_A_899_Mac_Mini_Replace_Your_200_AI_Subscription.md (auto-caption transcript; the YouTube title at listing time was “Don’t Buy a Mac Mini for AI Until You Watch This”)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=9imOOKYAvDM · Duration: 15:27 · Uploaded: 2026-10-23 · No sponsor named; a free model-per-Mac sheet is offered in the description.

Sharbel tests whether the 16 GB M4 Mac Mini that hosts his agents could also be their brain. His answer is no, not for hard work: the open models that fit are far weaker than paid frontier models, and running a frontier-class open model at home costs about $9,500 in hardware. The real reasons to run locally are privacy, no usage cap and near-zero running cost, which suit scheduled background jobs.

Key Takeaways

  • A subscription buys three things: the model, someone else’s data centre to do the reading, and an allowance before you are cut off. A Mac can in principle supply all three.
  • Memory decides what runs. The whole model must fit in RAM, and on a Mac that RAM is shared with macOS and apps. On his 16 GB machine, macOS and apps already used about 14 GB, and the GPU can use only about 10-12 GB. Models with “27B” in the name are 16-18 GB and do not fit.
  • Predict speed before downloading. Each token reads the whole model from memory, so tokens per second is roughly memory bandwidth divided by model size. His M4 (120 GB/s, per Apple) with a ~7 GB model should give 15-20 tokens a second; Qwen 3.5 9B measured 13. Run ollama ps: 100% GPU means the model fits; any CPU share means it spills and crawls.
  • Agents send long prompts. An agent sends its instructions, memory, tool list and history before each answer, and the model must read all of it first (prefill). A short summary prompt took 2 minutes 20 seconds locally against under 3 seconds for Claude. On Macs under 24 GB, Ollama defaults to about 4,000 tokens of context; an agent needs at least 64,000, so raise it in Ollama’s settings.
  • Waiting is fine for scheduled jobs. A morning brief that takes 10 minutes instead of 2 costs nothing while you sleep. The real limit is quality.
  • Quality gap. On the Artificial Analysis index he shows paid leaders around 58 and the best open models around 44-46 ^[ambiguous: Qwen 3.8 27B is recorded in this wiki at 52 on the same index]. The best open model is over a trillion parameters and needs about 236 GB even compressed.
  • The frontier-at-home maths. The largest Mac Mini has 64 GB (about 3,500 in Europe). Loading that model needs a 256 GB Mac Studio at about 200 a month, and is still slower and, after compression, less smart.
  • What local actually replaces is renting the same small model online, about 10 cents per million tokens, so a few dollars a month.
  • Three things no subscription offers: privacy (data never leaves the room), no cap, and running cost. At 4 W idle and 65 W flat out (Apple’s figures) and €0.26/kWh, that is about 74 cents a month idle and about €12 flat out all month.
  • What to buy, if anything. The new base M6 Mac Mini (1,300) is where bigger models fit; the 200 plan. The M5 Pro with 64 GB (from $2,700) reads memory about twice as fast. “32 gigs is the floor.”
  • Before buying, read a week of your own usage. If you are not hitting limits, drop to the 100 plan and keep the same models.

Try It

  1. Check your usage dashboard for a week; downgrade the plan if you never reach the cap.
  2. If you already have a Mac, install Ollama, pull a model about half your RAM size, and run ollama run <model> --verbose to see real tokens per second.
  3. Raise Ollama’s context length to at least 64,000 before pointing an agent at it.
  4. Route private and scheduled jobs to the local model and keep hard tasks on the paid one; Hermes and OpenClaw can hand a task off per job.

Open Questions

  • The 44-46 versus 58 index figures were read off a chart on screen; the index version and date are not stated, which may explain the gap with this wiki’s Qwen 3.8 27B figure.
  • Prices are as he quotes them for October 2026; regional prices vary.