Stability AI, a London-based tech startup focused on artificial-intelligence (AI) tools, has announced a new open-source large language model called StableLM. This large language model enables natural-language processing (NLP) researchers and developers to effectively generate text and code. It leverages the open-source Pile dataset, which holds Wikipedia, Stack Exchange, and PubMed information, allowing its models to be trained with more diverse data.
StableLM’s parameters are between 3 billion and 7 billion and the ability to handle 15 to 65 billion parameter models is being released in the near future. Its development is based on the mission of increasing the accessibility to AI tools and comes shortly after the company released Stable Diffusion – its image-to-text AI tool – via public demos, betas of the software, and a full-model download.
Stability AI’s initial language models succeeded due to their collaboration with the non-profit EleutherAI, but StableLM is the first to mark their foray into the open-source large language modeling space. Its systems will give developers access to pre-existing NLP tools so that they can better create customized solutions for their own projects. The software also intends to lower the bar to entry for NLP endeavors, by providing users a platform that can easily be transformed to suit their needs.
Stability AI was founded in 2018 by Yannick Assogba and Aimé Niyonkuru, both of whom have extensive experience in AI. Yannick Assogba is an AI researcher who has worked with the likes of MIT, DeepMind, and Twitter, while Aimé Niyonkuru brought years of software engineering knowledge to the table. The duo’s ambition is to share AI technology in an open source and ethical manner, while building a bridge between research and commercial applications. It has grown exponentially since its inception, developing various AI tools while forming strategic partnerships to increase its reach and impact. With the arrival of StableLM, Stability AI is furthering its commitment to creating tools built on transparency, robustness, and fairness.
