Editorial policy
Local AI Scope exists so that a technical decision — which AI model you can run on your own hardware, and whether you should — can be made by looking at figures rather than headlines. This page says who answers for what is published and by what rules it is decided.
1. Who answers for this
Editorial responsibility for everything published on this site rests with the person identified in the legal notice. There are no anonymous authors and no third-party contributions.
2. How content is produced and reviewed
There are two kinds of content here, and they are produced differently:
- The figures (sizes, context memory, per-language cost) are not typed by hand: they are generated from a versioned dataset and from the specifications each manufacturer publishes. An error in a figure is an error in the dataset, and it is fixed there, for all four languages at once.
- The prose is written and reviewed under human editorial control before it is published. No page goes live without a person having read it, checked its figures against the primary source, and taken responsibility for it.
That combination — real human review plus an identified person taking editorial responsibility — is also what Article 50 of the European Artificial Intelligence Act requires of anyone publishing information addressed to the public.
3. How “best” is decided
There is no editorial opinion about which model is best. There is a calculation, and the calculation is described in the methodology. In short:
- Work out what actually fits on the machine described, computing context memory from the attention pattern each model declares, layer by layer — not from the generic formula most calculators use, which can overestimate consumption more than fivefold.
- Apply the cost of the working language, measured with each model’s own tokenizer over an in-house parallel corpus.
- Apply the sensitivity policy: what information you will be handling, and whether it may leave your machine.
- Only at the end, and without any power to change the ordering above, look at commercial availability.
A result that cannot be explained by those steps is a bug, not an opinion.
4. Conflicts of interest
- No manufacturer pays to appear, nor for its position, nor for the tone of a page.
- The site sells no hardware, no software and no consulting.
- As of today there are no active affiliate links. The rules for when there are live in the affiliate disclosure.
- If hardware were ever lent by a manufacturer, it would be recorded on the relevant page and the measurement published regardless of the outcome.
5. What is not claimed
Saying what is not known is part of the editorial policy:
This site measures the cost of working in each language — how many tokens it takes to say the same thing — and does not measure the quality of a model’s answers in each language. That would require running the models and evaluating their output, which is not done today. No page on this site claims “measured linguistic quality”; when that measurement exists, it will be published with its protocol and its date.
The speeds shown on model pages are declared estimates derived from manufacturer specifications. They never decide on their own whether something fits, and they are labeled as estimates wherever they carry weight.
6. Corrections
A figure contradicted by a better measurement is an error to fix, not a position to defend. If you find one, write to the address in the legal notice giving the page and the source. Substantive corrections are published with their date; typographical fixes are simply made.
7. Sources
The primary source is always cited: the specification or configuration file published by whoever built the model, and the original license. This site does not host or redistribute model weights, and where it describes a license it links to the original text, which is the one that prevails.