AI21 Labs releases Jurassic-1

Israeli lab AI21 Labs launched AI21 Studio, a developer platform for building text applications, alongside the Jurassic-1 family of language models in open beta. The larger of the two, Jurassic-1 Jumbo, had 178 billion parameters — slightly more than GPT-3’s 175 billion — making it, by AI21’s account, the largest language model publicly available to developers at the time; a smaller companion model, J1-Large, had 7 billion.

The models used a 250,000-token vocabulary, around five times the size of GPT-3’s roughly 50,000, including multi-word tokens for common phrases and named entities. AI21 said the larger vocabulary improved tokenisation of varied text and reduced the number of tokens needed to represent a given piece of writing, lowering per-word API costs for customers. AI21 Studio also let developers create custom models from as few as 50–100 examples.

Jurassic-1 was not the first large language model to follow GPT-3, but it was among the first offered by a lab with no prior foundation-model product, distributed through a commercial API in direct competition with OpenAI’s own — an early sign that GPT-3-scale models were becoming a category other companies could enter rather than a capability unique to OpenAI. AI21 continued the Jurassic line through later versions before shifting toward smaller, more efficient Jamba models.