Jurassic-2

What Jurassic-2 is

Jurassic-2 is a family of large language models from AI21 Labs, released on 9 March 2023 as the successor to Jurassic-1^[1][2]. Like other language models of its generation, it works by continuing text: the user supplies a prompt, and the model produces what plausibly comes next, which can be an answer, a rewrite or a summary depending on how the prompt is framed. AI21 sold access to it as a hosted service, reached through a web playground or an API.

Sizes and instruction tuning

Jurassic-2 comes in three base sizes: Large, Grande and Jumbo^[1]. Alongside the base models, AI21 also shipped instruction-tuned versions of Jumbo and Grande that follow zero-shot instructions^[1]. Zero-shot means the user describes the task in ordinary words without providing worked examples; an instruction-tuned model is expected to carry out that request directly, instead of merely continuing the text in a plausible style. Offering both kinds of model lets a developer choose between a raw text generator and one already shaped to take directions.

Languages

Beyond English, Jurassic-2 supports Spanish, French, German, Portuguese, Italian and Dutch^[1]. Multilingual coverage of this kind matters for products that serve users in several countries, since one model can be reused across markets rather than needing a separate model per language.

Benchmark results

AI21 evaluated the family using the official repository of HELM, the evaluation framework from Stanford, and by that evaluation J2 Jumbo ranked second on Stanford's HELM benchmark^[1]. The company also said that Grande outperformed models up to 30 times its size^[1]. Both numbers come from AI21's own runs, so they describe how the company positioned its models rather than an independent ranking.

Task-specific APIs

Jurassic-2 launched together with a set of Task-Specific APIs, including a Wordtune API, covering paraphrasing, summarisation, grammar correction, text improvements and text segmentation^[1]. These endpoints package a single well-defined job rather than general text generation, so a developer who only needs, say, a paraphrase does not have to design a prompt and check the output format themselves.

Access and pricing

Developers could reach the models through the AI21 playground and API, free of charge until 1 May 2023^[1]. At launch AI21 also simplified its pricing to the total length of the text, with charges based on the total length of input plus output text^[1]. That replaced the more common split between prompt tokens and generated tokens with a single measure, which makes a call's cost easier to estimate in advance.

Place in the line

Jurassic-2 sat between AI21's first-generation model and the company's later work: it was succeeded in March 2024 by Jamba, which uses a hybrid Mamba-Transformer architecture^[2]. For roughly a year, Jurassic-2 was the model family through which most developers encountered AI21's language technology.