How to cite, and how to check

Three kinds of material, three rules. The same rules are stated on the terms and the data page.

A figure on a page or from the API

Individual figures shown on this site or returned by the API may be reused for any purpose, including commercially. Attribution is asked for, not required.

The downloadable dataset · CC BY 4.0

The downloadable dataset — the compilation, the computed columns and the evidence labelling — is published under CC BY 4.0. Reusing the dataset requires attribution.

Weights, model cards, configurations, product names

Model weights, model cards, configuration files, product names, trademarks and manufacturers' specifications belong to their owners and are not relicensed by this site; each is cited by its source URL and remains under its owner's terms.

Attribution for the current dataset version:

LLM Bottleneck open compatibility dataset (llmbottleneck.com/data), version 2026-10-03-13163f6c89cd, CC BY 4.0.

Whatever is reused, a computed or reconstructed figure must not be presented as a measurement: the evidence label is part of the figure.

For a model or device page:

LLM Bottleneck. "<page title>". Catalogue snapshot 2026-10-03. https://llmbottleneck.com/<path> (accessed <your date>).

For a measured accuracy figure:

LLM Bottleneck. "Accuracy Scorecard". Catalogue snapshot 2026-10-03. https://llmbottleneck.com/accuracy (accessed <your date>).

For the whole table rather than one figure, cite the open compatibility dataset by its version and snapshot: it is a fixed file with a published checksum, so a reviewer can confirm they are reading the same rows you read.

If you are citing an architecture figure — a layer count, a head count, a context length — cite the publisher, not us. We copied it from their config.json and the model page links the exact revision we copied it from. Their document is the source; this site is just a place it was read out loud.

1. Architecture. Every model page links its config.json at a pinned revision SHA, not at main. Open it. The layer count, head counts and context length on our page are that file's fields, unmodified.
2. Hardware. Every device field carries the manufacturer's own sentence and the URL it was read from, with the date. If the sentence is not on that page any more, the figure is stale and should be reported.
3. Weight sizes. Where a publisher released a GGUF, we use the file's own byte count from Hugging Face and show the file. Where they did not, the size is reconstructed and shown as a range; the accuracy page scores that reconstruction against every published file we do have, including the ones it gets worst.
4. Speed. Decode is calibrated against published benchmarks that are listed with their runtime build and commit. The calibration is leave-one-out validated and the corpus size is stated.

If a number here disagrees with the publisher's own document, the publisher is right and this is a bug worth reporting.

• Total runtime VRAM as a guarantee. The runtime reserve is an uncalibrated prior and every result says so on its face.
• Fine-tuning activation memory as a measurement. It is a model of what a framework keeps alive, and the fine-tuning page marks it separately from the terms that are arithmetic.
• Time to first token as a benchmark. It is a range with a physical floor and a utilisation solved from one published measurement on one device.
• A cluster speedup. We publish the interconnect ceiling, which is what a specification sheet supports, and not a speedup nobody has measured for your pair of devices.

The methodology page has the formulas, and the changelog records every correction, including the ones that were embarrassing.