Open supply fashions have change into a important a part of the AI panorama.
I used to be curious in regards to the tendencies within the open supply ecosystem, so I analyzed HuggingFace information on the highest 300 open supply fashions, each by general utilization & additionally the highest of the trending checklist.
Open supply fashions are ruled by open supply licenses. Much like common open supply software program, Apache & MIT dominate the licenses by mannequin rely. 76% of the highest fashions select one in every of these licenses. Apache is almost twice as widespread as MIT.
However the focus is larger when viewing the share by downloads. Fashions with Apache or MIT licenses symbolize 92% of downloaded fashions final month.
Stability, Fb, & Microsoft prime the creator checklist of open supply fashions by rely. So does TheBloke, an engineer who quantizes (or compresses) open supply fashions.
However the obtain information reveals very completely different patterns.
Meta’s fashions recorded 30% of downloads, pushed by its word2vec mannequin for speech recognition. Then OpenAI & Google not far behind.
The most well-liked fashions by downloads are fashions for coaching different fashions, known as Fill-Masks fashions. Then speech recognition. Third is textual content classification (LLMs are excellent at this.) Textual content technology is fifth.
How about reputation? HuggingFace likes of a mannequin are fully uncorrelated to downloads with an R^2 of 0.06.
Total, we are able to conclude extra lax licenses dominate the highest fashions. Meta, Google, Microsoft, Stability, & OpenAI are vital gamers inside the open supply ecosystem.
Speech is the most well-liked end-user utility of open supply fashions by downloads within the final month, outdated by testing – which is smart given what number of corporations are constructing or testing LLMs.
Given all of the innovation within the house, in 1 / 4 or two, this information may be very completely different. Who do you suppose will prime the charts on the finish of 2024?
