Jurassic-2 is AI21 Labs’ second generation of proprietary large language models, announced in March 2023 as a follow-up to the original Jurassic-1. It shipped in three sizes, Large, Grande, and Jumbo (later renamed Light, Mid, and Ultra), all trained on roughly 1.2 trillion tokens of web text, Wikipedia, books, and academic papers. AI21 never disclosed exact parameter counts, though the top-end model was understood to sit somewhere above 60 billion parameters. The line was aimed at businesses building on AI21 Studio, with task-specific APIs for summarization, paraphrasing, and text improvement layered on top of the general chat and completion endpoints. Jurassic-2 competed directly with GPT-3.5 at the time of release but has since been superseded by AI21’s Jamba models, which use a hybrid Mamba-Transformer architecture and much longer context windows.