A DECISION TOOL FOR AI BUILDERS
The home AI server question.
Should you buy a Mac mini or Mac Studio to run local models, or keep using frontier APIs? The useful answer is not a spec sheet. It is a workload, privacy, quality, and cost comparison.
Local models and frontier models earn different jobs. A home machine becomes worthwhile when a measured private or repetitive workload can repay its fixed cost. Until then, it may be an expensive way to avoid a calculation.
framework + calculator · no invented prices→ or scroll to advance · ← to go back
01 THE QUESTION BEHIND THE QUESTION
The box is one line on the bill
“Local is free” hides the purchase, power, storage, upkeep, and the time you spend keeping the thing useful. “Cloud is expensive” hides the fact that you only pay when you use it.
At home
With APIs
monthly frontier cost = model usage + hosted services
The hardware only has a payback story if both sides use the same workload and quality bar.
02 ROUTE BY JOB
Local and frontier are not substitutes
A local model is useful when privacy, repetition, latency, or offline access matters. A frontier model is useful when the task depends on broad capability, difficult synthesis, or a judgment you are still learning how to evaluate.
Local model
Cheap repetition
Run embeddings, classify or extract a large batch, search private notes, or keep a small tool available when the network is absent. The benefit is control over the path and the marginal cost.
Frontier model
Expensive judgment
Use a hosted model for hard synthesis, ambiguous reasoning, quality comparisons, or a task where the cost of a weak answer is higher than the API bill.
03 USE YOUR NUMBERS
The break-even math is small
No one else's electricity rate, call volume, hardware price, or token mix answers your question. Enter your own values. The calculator stays blank until you do.
Model the same job on both sides. If a local model needs human correction and the frontier result does not, put that correction time into local upkeep.
Your comparison
Enter all ten values to calculate.
04 ONE REAL BUILD
The home-server build answered a different question
The original question led to a shared context plane for Claude, Codex, and Gemini. It tested the memory boundary and the hand-off. The model-quality comparison came later, in Yawn.
Observed result
Raw transcripts stayed beside the source logs. Only filtered, re-redacted candidates crossed into the shared plane. Approval—not retrieval—remained the gate into future sessions.
Cutting Board
Its sessions tested a product boundary, not model quality: footage stayed local, and AI was described as ordering attention rather than deciding what is good. No local-versus-frontier benchmark was found there.
Yawn
Its evaluation tested the normal note-taking job against a hosted notetaker reference. The reference was useful for comparison, but it was not treated as ground truth.
05 THE MISSING COMPARISON
Yawn gave us the local-versus-hosted result
Two real meetings were scored by hand against the action items produced by a hosted notetaker. Local models ran through Ollama on Apple Silicon at temperature zero.
| Same calls | Hosted reference* | Local Llama 3.1 8B | Local Gemma 3 12B |
|---|---|---|---|
| Meeting A · shorter | 4 | 2 | 2 |
| Meeting B · longer | 6 | 1 | 2 |
| Total reference items | 10 | 3 | 4 |
*Reference is not truth. Every reference item was checked against the transcript first. The local totals describe omissions, not hallucination rates. The 8B and 12B models both produced true notes that omitted much of what the reference contained.
06 SAME AUDIO, DIFFERENT CHAINS
The best local result came from changing the pipeline
Meeting B was held constant while transcript source and speaker labels changed. The local end-to-end arm used local audio transcription, no speaker labels, and the same local note models.
| Arm | Transcript | Labels | Local Llama 3.1 8B | Local Gemma 3 12B |
|---|---|---|---|---|
| A | Hosted platform | Named | 1/6 | 2/6 |
| B | Hosted platform | Stripped | 0/6 | 3/6 |
| C | Our local audio | None | 0/6 | 4/6 |
The local transcript preserved every term needed for the six scored commitments, used 91% of the hosted transcript's word count, and ran at 25× realtime on a laptop. This was one meeting and one configuration, not a general law.
07 THE DESIGN TURN
The Mac mini became optional
Once the shared memory system ran on an always-on public endpoint, the home server stopped being the center of the architecture. It became a possible worker.
Shared plane
Cloud services did the job
Authenticated ingress, durable queueing, reviewed search, audit history, and portable snapshots were enough for the shared layer.
Private work
Each machine kept its evidence
Transcript parsing, local SQLite mining, redaction, and the durable outbox stayed beside the raw source logs.
Routing seam
Frontier remained a choice
A gateway can route hard work to hosted models, but model distillation stays disabled until quality, privacy, and cost are measured.
Honest use for a box
Buy only for a measured gap
A home machine earns its place for private embeddings, large repeated batches, or offline access. Not because the diagram looks better with one.
08 WHAT THE LOCAL STUDY FOUND NEXT
Bigger was not the answer
On a real capture ledger, model scale and prompt batching changed the result in surprising ways. The useful improvement came from making one candidate decision per call, then pruning locally.
| Local validation cell | Kept before prune | Kept after prune | Recall |
|---|---|---|---|
| Meeting 1 · Ollama | 71/165 | 35 | 12/13 |
| Meeting 1 · MLX | 104/165 | 27 | 12/13 |
| Meeting 2 · Ollama | 90/300 | 48 | 13/13 |
| Meeting 2 · MLX | 152/300 | 54 | 13/13 |
The four-cell result used Gemma 3 12B, single-candidate calls, and a deterministic contiguous-run collapse. It is a measured local pipeline result, not a claim that Gemma 3 12B is a frontier replacement.
09 THE COST THAT WAS ACTUALLY MEASURED
Local cost showed up as bytes, memory, and time
The sessions measured operating constraints. They did not produce a trustworthy dollar comparison yet.
Measured locally
One installed local note pass produced the useful overview/decisions/follow-ups shape in about 28 seconds. MLX admission runs also recorded model load, call time, resident memory, model-tree digests, and strict response failures.
Still missing
No verified Mac mini or Mac Studio purchase price, long-run wattage, API invoice, token bill, or correction-time ledger for the same workload. Those are the numbers the calculator asks you to supply.
10 THE LIMITS
The evidence is useful because it stays narrow
We have a measured local-versus-hosted note-taking result. We still do not have a complete home-server business case.
Not measured
Actual electricity draw and maintenance time for a Mac mini or Mac Studio under a repeatable workload.
Not produced
No finished Mac Studio-versus-frontier monthly TCO spreadsheet. The calculator here intentionally asks for the missing inputs.
Not general
Yawn's comparison used a small set of real meetings and a hosted reference that may also omit commitments. It does not establish universal model rankings.
What is supported
Local models can be useful for private, repeated, offline work. Frontier models can be worth their cost for hard judgment. The boundary should be measured per workload.
11 START SMALL
The first purchase is a measurement
If you are curious about a home AI server, you can answer most of the question before buying one.