Usage statistics of MTEB (2025/12/03)

[KennethEnevoldsen]

So looked a bit at some download (usage?) statistics of MTEB and found that they were worth sharing. In general, I would use the opportunity to thank the MTEB maintainers and contributors, great work!

Do more people use MTEB?

Overall usage has grown slowly but consistently, you can see some notable spikes - I suspect those co-occur with releases and potentially some social media posts:

Does that have any relation to GitHub stars? (you know, the only real number that matters ;) )

It has! Although growth seems fairly steady

What do people use with MTEB?

Impressively, it seems like almost all users are on Python 3.12 and Linux - way to go users!

What version do people use?

People still seem to be on v1:

Limitations

This does not include the influence of the leaderboard, which probably has a bigger reach compared to the package and likely includes downloads from bots, CI etc.

Edit: Links: Download stats and star history, version specific downloads, clickpy

---

Any cool stats that I am missing out on? Do feel free to share

[KennethEnevoldsen]

Here is a small overview of the current status on the leaderboard (sadly we can't see it across time):

and 6,766 likes (hearts)

[Samoed]

Interesting, why almost 99% Linux users?

[KennethEnevoldsen]

I suspect WSL, people using servers (I work on macos as well as 2-4 linux servers) then of course automated workflows and bots that keep a copy of the pypi.

[tomaarsen]

It's very common to have an extremely high % of Linux users. Even projects that are certainly used by Windows etc. users like numpy or nltk show 90%+ Linux.

[tomaarsen]

you can see some notable spikes - I suspect those co-occur with releases and potentially some social media posts

I strongly suspect that there are simply from automated downloads. The Python version graphs are a good indicator here: if only one version spiked, then it was presumably automated.

It's good to see the steady growth inbetween the outliers, and slowly you're starting to see the weekdays vs the weekends. I always feel like this is a good example of real users, as they'll mostly use projects during the weekdays. See e.g. the nltk stats for a good example of what that'll look like.

P.s. I also like the clickpy page that you linked (https://clickpy.clickhouse.com/dashboard/mteb), I feel like this page has some valuable stats that no other big tracking sites do.

• Tom Aarsen

[KennethEnevoldsen]

Ahh good point with the weekly activity - Yeah didn't know about clickpy (Stephan shared it), but I agree it def. had stats that I couldn't find elsewhere

[KennethEnevoldsen]

Someone shared this with me as well:

from the docs:

The resolution time is the median time an issue or pull request stays open. Some issues are ignored based on their labels (example of ignored labels: "enhancement", "feature"…). Issues older than 6 months are also ignored.

The percentage of open issues and pull requests. Some issues are ignored based on their labels (example of ignored labels: "enhancement", "feature"…).

Which I think is decently close to what we want.

source: https://isitmaintained.com/