Skip to content

Trending AI Today

How the edition is made

Method · sources, the one rule, the schedule

Two public lists, one arithmetic rule, one record per day. This page states all of it, with the constants the code uses.

Sources

The collector reads two addresses of the public Hugging Face API, without a key. Models come from the model list sorted by trending score; the first 40 are kept. Papers come from the daily papers list of the current day; the 30 most upvoted are kept, the first 12 with the complete abstract.

  • https://huggingface.co/api/models?sort=trendingScore&limit=60
  • https://huggingface.co/api/daily_papers?date=YYYY-MM-DD&limit=100

When the current day has no paper list, which happens on weekends and holidays, the collector looks back up to 4 days and uses the most recent one. The edition then says which day its papers belong to. For the leading models that are not themselves quantisations, one extra request each finds the three most downloaded repositories that declare the model as their base.

Sources named in earlier plans for this site were dropped after checking them: Papers with Code now redirects to Hugging Face Papers, and Product Hunt requires a token and restricts commercial use of its data. GitHub repositories are left to GitHub Trending Today (opens in a new tab).

The one rule

The order of models is the order Hugging Face returns. A model is moved out of the ranking, into a list of its own, when all three conditions hold:

  1. it has 100 or more likes;
  2. the repository was created 3 or more days before the collection, because download counts lag for new repositories;
  3. its downloads over the last 30 days are fewer than its likes divided by 10.

The reason is that the trending score responds to likes, and likes can be gathered without anyone using a model. On the day this rule was written, the first model of the trending list had about five thousand likes and zero downloads. The rule does not say why; it reports the two numbers and steps the entry aside.

What is never done

  • No score of our own is computed, and no model is promoted.
  • No abstract is summarised, translated or rewritten. Excerpts are cut at a word boundary and marked.
  • No field is guessed. A missing parameter count, licence or code link is shown as missing.
  • No thumbnail is generated for a paper, and figures from papers are not copied.

Schedule and storage

The collector runs on a timer every three hours, starting at 00:07 UTC, so the current day is refreshed as votes and downloads move. Each run writes the latest edition and the day’s compact record: the top 15 models and papers with their numbers. The last run of a day is the one the archive keeps.

If the model list cannot be read or comes back empty, nothing is written and the previous edition stays, with its date. Every page prints the collection time, and shows a notice when the edition is more than 36 hours old.

The snapshot shipped with the site

The pages you receive as HTML are built from a snapshot collected when the site was built, by the same code the scheduled collector runs. That snapshot is labelled “shipped with the site build” wherever its numbers appear. When the collector has a newer edition, your browser replaces the lists on the home, models and papers pages and the label changes.

Limits you should know

  • Download counts are Hugging Face’s 30-day figures. Models served mainly through an API or another host look smaller than they are.
  • Licence tags are chosen by uploaders and can be wrong or differ from the base model’s licence.
  • Paper upvotes measure the attention of one community on one day.
  • Day-over-day changes need two editions; a model’s first day on the list shows none.

Source: Hugging Face public API · updated

Read them yourself

Primary sources

Questions readers ask

Do you use the Hugging Face API with a key?

No key is required for the two addresses we read, and none is used. The collector identifies itself in the User-Agent header and makes at most a few dozen requests per run.

Can the held-back rule be wrong?

Yes. A model served mainly through an API, a gated repository or another host can have many likes and few Hugging Face downloads for honest reasons. That is why held-back models stay visible with their numbers instead of being deleted.

Why not include GitHub repositories and news?

GitHub repositories are covered by GitHub Trending Today, a separate site in the same family, and we link to it from papers that have code. News about releases is a different job from reading two public lists; this site does only the second.

What happens when Hugging Face is unreachable?

If the model list cannot be read or comes back empty, the run writes nothing and the last good edition remains with its date. If only the papers request fails, the previous paper list is kept and the run records an alert.