Introducing J2

AI21 Studio unveiled Jurassic-2(J2) and Task-Specific APIs, delivering reading and writing AI in a form that's easy to attach to real products.

J2 is a next-generation foundation model that broadly improves upon the existing Jurassic-1. It offers 3 base models—Large, Grande, and Jumbo—plus instruction-tuned models for Grande and Jumbo, emphasizing zero-shot instruction-following, lower latency, and multilingual support.

The key improvements are as follows.

• Quality: In internal HELM evaluations, J2 Jumbo recorded an 86.8% win-rate.
• Cost and speed: Grande showed performance on par with models up to 30x larger, and overall latency was said to be up to 30% faster than the previous model.
• Multilingual: Supports Spanish, French, German, Portuguese, Italian, and Dutch.
• Instruction-following: The model can be controlled through natural language instructions alone, without examples.

At the time, J2 was offered free of charge until May 1, 2023, and pricing was simplified based on total text length combining input and output. It was immediately available in the Playground and via API, along with tips for using the instruct models.

Under Task-Specific APIs, the Wordtune API set provided Paraphrase, Summarize, Grammatical Error Correction(GEC), Text Improvements, and Text Segmentation. AI21 focused on reducing the process of fine-tuning general-purpose LLMs directly, and instead making frequently repeated reading and writing tasks immediately usable.

Benchmark results were also highlighted. The Summarize API was claimed to have 19% higher faithfulness than OpenAI's Davinci-003, along with an 18% higher human acceptance rate. The Paraphrase API was said to have latency about 1/3 that of OpenAI's, with a 33% edge in diversity and an 8% edge in meaning preservation.