AI21 Labs Simplifies Jurassic-2 Offering

Three months after launching Jurassic-2 (J2), AI21 Labs simplified its offering based on user feedback. Instead of the previous 5 foundation models, only 3 models remain, and all three now support instruct capabilities, allowing a single model to handle both zero-shot prompting and few-shot prompting.

This decision is based on internal testing showing that instruct models performed equal to or better than non-instruct models on Stanford's HELM and various few-shot datasets. The model names also changed from Large, Grande, Jumbo to Light, Mid, Ultra, making their size and role clearer.

• Ultra: The largest and most powerful model, suited for complex generation and comprehension tasks. It recorded a win-rate of 86.8% in internal HELM evaluations, but also has the highest cost and latency.
• Mid: A mid-tier model aimed at balancing quality and cost, suited for question answering, summarization, copy generation, advanced information extraction, and more.
• Light: The smallest, fastest, and most cost-efficient model, tailored for keyword extraction, sentence classification, NER, short-form copy generation, sentiment analysis, and more.

AI21 explained that this reorganization allows users to choose the right model more quickly. AI21 Studio users do not need to take any immediate action, and API changes will be handled via automatic rerouting.