About this site

Downloading a local model is a commitment of 5 to 40 GB made on the strength of a forum post. The advice you find is either a rule of thumb that ignores context length, or a table that treats every model of the same size as identical. Neither survives contact with a real machine, and both cost people an evening and a lot of bandwidth.

So this site does the arithmetic properly. It reads each model's own architecture, meaning layer count, attention design, expert routing and cache layout, and works out what that specific model needs on specific hardware at the context length you intend to use. It covers 86 language model architectures, 131 hardware profiles, and separate engines for image, video and audio models. Every formula is written out in full, and the engine is checked against a real machine on the validation page.

Who runs it

Stéphane Cardon is a network and cybersecurity engineer. He works on network architecture, system administration and security, and runs local models on his own hardware, which is a modest RTX 3050 with 6 GB of VRAM. That card is the reason the site cares about what actually fits. It is also the machine behind the measurements we publish, and it is the reason the small-card guides were written by someone who has hit the wall on one. You can find him on LinkedIn, and the guides carry his name as author.

How the content is made

We use AI tools as part of the work, and we would rather say so than leave it for you to wonder. They help write the code and draft and edit the articles. They do not produce the numbers. Every figure in a guide, such as the memory a model needs, the speed on a card or the size of a cache, is computed by the site's own engine at build time, so the text cannot drift from the calculator. Commands and flags are checked against the official documentation of the runtime they belong to, and the first-hand measurements come from the machine described on the validation page.

Stéphane Cardon directs the project and is accountable for what is published. If you find an error, the correction policy below is how it gets fixed. We hold the articles to a plain standard: say what the number is, where it came from, and how sure we are of it.

How it is paid for

Advertising, through Google AdSense, and Amazon affiliate links on the hardware pages. If you buy a card through one of those links we earn a commission and you pay the same price. We do not take payment for placement, no vendor has any say in a ranking, and the buying tables are ordered by memory, bandwidth and a published formula, not by what pays best. The privacy page explains what the ad and affiliate services collect and the choice you have over it.

Where a recommendation would be bad advice, we say so even when the alternative is a cheaper purchase. The tables refuse to recommend a quantization low enough to break a model, which routinely means telling someone their existing card is fine and not pointing them at a bigger one.

Corrections and updates

Each guide shows when it was published and when it was last updated, and an update means something changed: a number, a recommendation or a command. Changes to the calculation engine, the catalog or the price data go in the update log. Prices are reviewed by hand and show the date of the last review. When we find that an earlier statement was wrong we fix it in place and note the change in the log, and we do not quietly rewrite history.

Measured numbers that disagree with ours are the most useful thing a reader can send, since that is how the estimates get better. Missing hardware and missing models are worth reporting too. The contact page has the details.

What we do not do

• No accounts, no sign-up, no newsletter.
• No tracking of what you check. The calculator runs in your browser, and your hardware choice is stored locally so you do not have to pick it twice. It is never sent anywhere.
• No quality claims we cannot back. We publish estimates and label them as estimates, and we say which constants we measured.
• No model leaderboards. Plenty of places argue about which model is smarter. We answer the narrower question of whether it fits and how fast it runs.