Is Meta's Open Source AI Bet Paying Off?
OpenAI's pre-ChatGPT large language model (LLM) GPT2 released in 2019 was the most downloaded text generation model on the LLM repository HuggingFace with 15.5 million downloads in the past month. Unlike many more modern models, the training data only encompasses "all the web pages from outbound links on Reddit which received at least 3 karma" and excludes all Wikipedia content according to its model page on HuggingFace. Meta's AI ventures are also present in this listing, although not with their most recently released models.
OPT-125M, released by Meta in the summer of 2022 but still attributed to Facebook on HuggingFace, saw 6 million downloads in that timeframe. Meta's older flagship model Llama 3.1 still comes third overall with 5.8 million downloads. Since our chart only displays the most recent model iteration, Llama is absent from the top 5, with its most recent version, Llama 3.3, accumulating 597,000 downloads in the past month. Other entrants in the top 5 downloaded models in the past month apart from OpenAI and Meta are MistralAI's Nemo Instruct model, which was downloaded 1.5 million times and Apple's OpenELM 1.1B Instruct model, which racked up 1.4 million downloads.
As opposed to proprietary applications like ChatGPT, open source models can in theory be downloaded by everyone and used, adapted and modified according to the respective license they are released under. However, even with freely available models like Llama, it's often unclear what data they are trained on, and usage can be complicated, requiring additional software and knowledge.