If you want to try a new generative AI model, you have several options. Their developers may offer their own web service – as do OpenAI, Google, Microsoft, Meta or Anthropic – but in many cases, especially for Open Source developers, this is not the case. And that is where Hugging Face has become an absolute reference.
No downloads and installations. At Hugging Face they have managed to significantly facilitate access to these models. The platform, with a structure that is slightly reminiscent of GitHub, brings together all types of models so that if we want we can “deploy” them and use them as if we were doing it on our own PC.
The number of models has increased dramatically in recent months. Source: Hugging Face.
More than a million AI models. As they point out in Ars Technica, the platform already has more than a million AI models available, something made clear by the immense number of projects that have emerged since ChatGPT began that explosion of generative AI.
Started as a chatbot. It is curious, but in its beginnings in 2016 Hugging Face began its journey as a limited AI chatbot. In 2020, two years before the launch of ChatGPT, they pivoted and became a hub of AI models with an Open Source philosophy. The number of models on the platform has grown dramatically, especially since the beginning of 2023. Last year, however, they launched their own Open Source rival to ChatGPT.
Testing the new Llama 3.2 (3B) thanks to Hugging Face. As expected, it still makes a mess counting R’s, as happens in most models. Even so, the important thing is that being able to test models like this with a click is very easy thanks to this platform.
Great models, but also small. Clement Delangue, co-founder and CEO of the company, celebrated the milestone by highlighting that the platform offers access to large models (Llama, Gemma, Phi, Flux, Mistral, Stable Diffusion, Grok, Whisper, among them), but “also to another 999,984 “.
Long live specialized models. For Delangue, the message of “one model to dominate them all” is a fallacy. According to him, “it is better to have smaller models, customized and optimized for your use case, your domain, your language, your hardware and, in general, your limitations.”
Forks everywhere. This work of “adjustment” and customization of models has driven this growth in available models. Many start from the same base and then undergo modifications and “fine-tuning” processes from which derived models appear, something like software project forks. Llama, Meta’s LLM, is a good example of this: many models derived and adjusted for specific use cases have come out of it.
Of everything and for everyone. At Hugging Face we can find models for all tastes. We have traditional text chatbots and variants dedicated to processing text, but also specialized models for image classification or object detection. Its list of the most downloaded and used is a demonstration of that variety: the first is a model for classifying audio content, while the second, BERT (from Google) is used in language modeling projects.
In WorldOfSoftware | The European AI Law has just been accepted by more than 100 companies. There are two big absences: Apple and Meta
