Skip to main content
Every model Kyma serves can be put beside any other on a URL. Sort two to four model ids alphabetically, join them with -vs-, and that is the page:
Any other ordering of the same models permanently redirects to the sorted one. One comparison has one URL.

​
What is on the page, and what is not

Three things, and nothing else: There are no benchmark scores on these pages. Kyma publishes what it measures and what a creator publishes about its own model; a third-party scoreboard measuring somebody else is neither, it goes stale without telling anyone, and a page carrying it would be making a claim it cannot refresh. So a comparison page will tell you what a model costs and whether it answers, and it will not tell you which model is smarter. Availability is the same number, computed the same way, as the one on /models and /status: successful observations divided by total observations over 30 days, counting Kyma’s scheduled probe and real customer requests alike. Throughput and latency are probe-only, from one identical prompt sent to every model every six hours, so the only variable left between two rows is the model.

​
Which comparisons get indexed

Every combination renders, and every combination has a canonical URL. Not every combination is submitted to a search engine, because a site that asks to be indexed on tens of thousands of near-identical pages teaches a crawler that most of it is not worth fetching. A curated comparison is always indexed. Anything else is indexed when both of these hold:
  1. Every model in it has at least 100 observations in the last 30 days. A model measured since yesterday has a real reading on a sample thin enough that one failure moves it by a point, so a comparison resting on it is not something to rank.
  2. It is one of the curated comparisons, or it has been opened on at least three separate days. Curated comparisons are the matchups a buyer actually shops between, listed on /compare. Anything else has to have been asked for on three different days, so three reloads in a row do not count as three occasions.
Everything else is noindex, follow: the page works, it renders, it shares, and its links are still crawled. Only the request to rank it is withheld. Column count is not part of the rule. A four-model comparison people open is indexed on the same terms as a two-model one.

​
Finding comparisons programmatically

/compare lists the curated set, the most-opened combinations among those opened in the last 30 days, and the last 20 opened. views is a lifetime total: window chooses which combinations are listed, it does not window the counter. The same data is available directly:
No authentication, cached for five minutes, and rate limited to 60 requests per minute per IP. sort is views or recent, window is 7d, 30d, 90d or all. seen_days is the number of distinct days the combination has been opened on, and it is what the index rule above is measured against. The counter behind it records a slug, the model ids in it, two counts and three dates. It holds no user id, no IP address, no session and no referrer.