How we score

The Local AI Reality Check is deliberately simple — and deliberately honest about where the numbers end and the estimates begin.

The core idea, in one line

Memory decides what fits. Bandwidth decides how fast. We keep those two verdicts separate, and in the current version we only ship the first one.

How fit is calculated

Every Apple Silicon Mac has a fixed amount of unified memory shared by the CPU and GPU. A model's size — its parameter count times its quantisation level — determines how much of that memory it needs. Fit is arithmetic against the machine's published memory ceiling.

Concretely, a model is said to fit when its required memory (its minRamGb in our dataset) is at or below your Mac's unified memory. We reserve roughly 6GB for macOS and a browser before counting models.

Fit is an estimate, not a measurement. The minRamGb figures come from a CC BY 4.0 dataset by ModelFit, which derives them from a RAM-budget ratio (about 0.7 of total memory below 32GB, scaling to 0.85 at 128GB and above). We surface them as the source's estimates and label every fit result accordingly.

Why we don't ask for your GPU core count

You know your memory. Almost nobody knows their bandwidth or their GPU core bin. For every Max-class chip, the memory configuration maps to exactly one bandwidth figure — so we derive bandwidth from memory and never ask for it. On the M6 Mac mini, memory alone determines bandwidth (16GB = 153 GB/s, 24/32GB = 170 GB/s).

The verdict scale

VerdictWhat it means
Too smallThe machine can't do the job at usable quality. We say so plainly — this is the verdict that makes the others believable.
FitsThe machine does the stated job.
Fits with roomIt does the job plus headroom — and we explain what that headroom buys.
Beyond your stated needIt clears the job with capacity left over. We never frame this as a mistake — headroom is options, not waste.

The rules we refuse to break

Never over-recommend. Every recommendation is gated by the machine's memory ceiling. We will not suggest a 70B Q8 build to a 32GB machine — the math forbids it and we let it.

Some models fit no Mac that exists. Three models in our catalog need more unified memory than Apple has ever shipped — Llama 3.1 405B and Llama 4 Maverick want 320GB (a tier that has never existed), and DeepSeek-R1 671B wants 512GB (the announced, not-yet-orderable M5 Ultra). We return an honest “no Mac fits” for these rather than recommend the impossible.

Privacy and compliance change the question. If your reason is data that can't leave your machine, local isn't the cheaper option — it's the only option. We suppress break-even entirely and answer feasibility instead.

MoE models are not dense models. A mixture-of-experts model keeps every expert resident in memory (fit uses total parameters), but only activates a fraction at a time (speed uses active parameters). We carry that flag in the data even though speed isn't scored yet.

What we don't score yet

Speed (tokens per second). A credible speed number needs measured benchmarks, and today we have only a couple of calibration points — not enough. Shipping a speed figure now would mean shipping it wrong. Fit ships with confidence; speed arrives in v1.1 as a labelled band, and any figure above 128GB will be explicitly marked as extrapolated.

Precise context-window sizing. KV-cache memory is only populated for a minority of the catalog, so we don't present per-model context estimates we can't back.

Break-even method

Months-to-payback divides an approximate starting MSRP for the machine's tier by your current monthly cloud spend. Prices are coarse tier estimates (base ≈ $799, Pro ≈ $1,999, Max ≈ $2,999, Ultra ≈ $4,999), not quotes — price your exact config before deciding. And the honest caveat: below roughly $800/month of cloud spend, local rarely wins on cost alone.

Sources

One known gap: M1 base bandwidth is unpublished by Apple (it's only stated relative to M1 Pro). We mark it “unpublished” rather than present a derived number as an Apple figure.

← Back to the calculator