HomeGuides

Guides

Guides that answer with a measurement

Every guide here starts from a figure measured on this site that contradicts something widely repeated. The table says what each one settles and the number it brings, so you can tell which is worth your time before opening it.

The guides Measured
Guides published6
In English3
Silos covered2
Most recent2026-08-25
Guides published in English: the silo of each one, what it settles and the figure it measures.
GuideWhat it settlesThe measured figure
ChooseOn German code, qwen3-coder-30b is tied last of the 18 models we measuredA coding model picked on English benchmarks can be the most expensive one on a German codebase: qwen3-coder-30b needs 353 tokens where English needs 210.353 vs 210
PrivacyLLM token calculator: what the same text costs in four languagesWhat a language costs depends on the model, not only on the language: the same document is ×1.62 tokens in German on qwen3-8b and ×1.29 on gpt-oss-20b.×1.62
ChooseLLM VRAM calculator: what really fits, counted layer by layerThe generic VRAM formula treats every layer as full attention. On gemma4-12b at 32k it claims 12.00 GiB of cache where the real figure is 2.31 GiB.×5.2

Every figure in this table is measured by Local AI Scope on the real published files and a parallel corpus, and each guide carries the date of its measurement. None of them is a vendor claim.

Other languages

The same method, measuring another language

These are not translations of one another: each one counts what its own language costs on the same 18 models. They are here because the answer usually lives in the language you actually work in.

Guides published in the other languages, each one measuring its own language.
GuideWhat it settlesThe measured figure
Auswahl · DeutschPC für KI-Anwendungen: was auf Deutsch wirklich hineinpasstDie Sprache ändert, wie viele Modelle auf dieselbe Karte passen: auf einer 12-GB-Karte mit einem Vertrag von 12.000 Wörtern 6 auf Englisch und 3 auf Deutsch.6 → 3
Choisir · FrançaisQuelle carte graphique pour une IA en local : ce qui tient vraiment, en françaisLa langue change le nombre de modèles qui tiennent sur la même carte : sur une RTX 4060 de 8 Go avec un contrat de 12 000 mots, 4 en anglais et 3 en français.4 → 3
Elegir · EspañolQué PC necesitas para IA local: lo que cabe de verdad, en españolEl idioma cambia cuántos modelos caben en la misma tarjeta: en una RTX 4060 de 8 GB con un contrato de 12.000 palabras caben 4 en inglés y 3 en español.4 → 3

A guide tells you what happens in general. What fits on your machine — with the context your work needs, and in your language — is what the Model Finder answers.

Check it against my machine