Athomic Smithy
Contents
Plate 00

Local inference, kept alive on purpose

From a cold rig to a paid product, with every number measured.

Athomic Smithy is one playbook and the scripts that go with it: how to run a local model 24/7 on a single Apple Silicon Mac, and sell something from it.

The chapters are the parts nobody else finishes — watchdog, tunnel, queueing, fallback, cost arithmetic. Every recipe names the rig, the software versions and the numbers it was validated on.

It does not promise a single dollar of passive income.

Exploded line drawing of the reference rig: the enclosure, the logic board and the heatsink, with a single traced path leaving the machine through a tunnel to three clients.
Fig. 1 — Reference rig, exploded, and the one path that carries a request off the machine. A drawing, not a photograph.
Reference rig
Mac mini M4, 16 GB
Second tier
M4 Pro, 24–48 GB
Playbook
$47
Playbook + Bench
$197
§ 01The problem

The model is not what stops you.

A developer looks at an idle M-series Mac with more unified memory than most cloud instances, and knows local models are now good enough. Then the questions start. The answers are scattered across walkthroughs, forum threads and vendor docs.

  1. 01Which quantised model actually fits 16 GB and still answers usefully.
  2. 02How to serve it behind a stable OpenAI-compatible endpoint.
  3. 03How to reach a home machine from the internet without opening the router.
  4. 04How to keep it alive through reboots, thermal throttling, macOS updates and sleep.
  5. 05How to queue requests when two users arrive in the same second.
  6. 06What to do when the machine is genuinely down and paying customers are waiting.
  7. 07What a thousand requests cost on your own rig, against the API bill.

Nothing joins those dots from a cold machine to a product with paying users. So the rig stays idle, or the stack gets assembled over three weekends and abandoned after the first 3 a.m. outage.

§ 02The arithmetic

Four things to do with a Mac that has spare memory.

One of them is being sold loudly right now. It is the weakest of the four, and the numbers are public.

Rent the rig's idle compute to a token network

Reported by real providers on Apple Silicon networks, against calculators promising $75, at single-digit utilisation.

Around $18 per month

Keep paying per token to a hosted API

The margin on everything you ship goes to the model vendor. None of it accrues to hardware you already own.

Your bill, forever

Rent cloud GPUs, or rent a hosted Mac mini

rentamac.io from $3.30 per day, myremotemac.com from $85 per month. Both reintroduce the monthly bill you were trying to kill.

From $3.30 per day

Sell a product served from hardware you own

The marginal cost of a request is power. The playbook shows that arithmetic rather than asserting it.

Electricity only

Sourced from public provider reports and vendor pricing pages, not run on our rig. Figures as of 2026-08-28.

§ 03The stamp

Every recipe ends with a line like this, or it does not publish.

The validation stamp names the rig, the memory, the operating system, the server version and the date it was last run. When something was read in a vendor document instead of run here, the recipe says exactly that.

Specimen — the format, not a published measurement

Validated on Mac mini M4, 16 GB, macOS 15.3, Ollama 0.5.7 — 2026-04-02

Rig

The exact machine and its unified memory.

Versions

Verbatim and complete, never approximate.

Date

A revision, re-run when the software moves.

§ 04Contents

One spine, in the order you will need it.

Nine recipes carry a rig from a sealed box to something a stranger can pay for. Each one names the machine it was run on.

  1. 01Choose the modelBy task and memory budget, per rig tier.
  2. 02Serve the endpointA stable OpenAI-compatible surface on the rig.
  3. 03Expose the rig without opening a portA tunnel, never port forwarding.
  4. 04Keep the endpoint alive after a rebootlaunchd, a watchdog and a restart policy.
  5. 05Queue and cap concurrencyOne accelerator, many callers.
  6. 06See the rig get sick firstObservability before customers notice.
  7. 07Degrade instead of dyingA documented cloud fallback for outages.
  8. 08Do the unit-cost arithmeticCost per million tokens against current API pricing.
  9. 09Pick a product shape that worksHigh volume, low glamour, privacy-bound work.
§ 05The two packs

Two objects. One price each, paid once.

Playbook

$47
  • The playbook as a PDF, plus the scripts and config files.
  • Lifetime access to every later revision.
  • A ten-message taste of the Bench.

Playbook + Bench

$197
  • Everything in the Playbook, plus the extended recipe set.
  • Twelve months of the Bench, the assistant that answers from the playbook.
  • A fair-use ceiling of 60 messages per day, published so inference stays predictable.

Not included

  • No Intel Mac coverage.
  • No Windows or Linux.
  • No promise of revenue.
  • No one-to-one support.
§ 06Status

The first revision is being validated on the reference rig.

Nothing is for sale on this page yet, and there is no form here to join anything. When a recipe has been run end to end on the rig named in it, it publishes, and the two packs open.

Until then, the honest thing is to say so. Keep the address: the recipe index and the cost calculator land at the same place.

Copying the address is all this button does. Nothing is charged, nothing is collected.

Revision
2026-08-28
Recipes published
0
Packs on sale
0 of 2