How this compendium is built

What this is

Global Distillation is a compendium of one technique: training a small student model on the behaviour of a large teacher. The same technique is a research method, a product line, a pricing strategy, a contract term and, since 2025, an accusation. Each of those readings gets its own section, built from the same catalogue of sources.

The site is static. Perspective research is written into JSON files that follow a published schema; the browser fetches them and renders every table and chart from the same data a reader can download. There is no server-side model, no summarisation step between the source and the figure, and no number that exists only in a chart.

Where the numbers come from

Sources are ranked in this order: the paper or technical report; the vendor’s own model card, pricing page or terms of service; a filing, statute or official memorandum; then reporting by a named outlet. A claim that exists only in reporting is attributed to whoever made it, in the text, not presented as fact.

Prices are list prices for the standard tier on the date shown, in US dollars per million tokens, excluding batch, cache and volume discounts, because those vary by contract. Benchmark scores are the figures the model’s own authors published, which is a real limitation: labs choose their comparisons. Where an independent reproduction exists, both are shown.

Download counts, star counts and paper counts are activity measures collected from public APIs. They measure attention, not quality, and they are labelled that way wherever they appear.

What this method cannot tell you

Three of the most interesting questions are unanswerable from public evidence, and this site does not pretend otherwise: whether a given closed small model was distilled, what a training run really cost, and whether a particular model was trained on a competitor’s outputs. Evidence offered in public disputes is circumstantial and is reported here as a claim by a named party, with its rebuttal.

Benchmark retention percentages compare a student against its own teacher on one benchmark. They do not transfer across benchmarks, and they say nothing about robustness, long-context behaviour, or how a model degrades outside the distribution it was distilled on.