Israeli startup AI21 Labs has released a large-scale language model (LLM) called Jurassic-2, VentureBeat reported on the ninth (local time).

With a vision to make artificial intelligence (AI) a human thought partner, AI21 Labs is making a next-generation language model that developers and businesses can customize for specific tasks and construct text-based applications faster in multiple languages. Disclosed.

An upgrade of the Jurassic-1 model released in 2021, Jurassic-2 is largely a language model that will be used to reply questions, write poetry and essays, and summarize texts.

In comparison with Jurassic-1, the response time is as much as 30% faster, and it incorporates a prompt adjustment function that permits accurate results to be obtained. It supports six languages ​​besides English: Spanish, French, German, Portuguese, Italian, and Dutch.

Jurassic-2 is an element of AI21 Studio (AL21), a natural language processing (NLP-as-a-Service) platform that permits developers to create text-based applications resembling virtual assistants, chatbots, text simplification, content moderation, and artistic writing. Studio)’.

To this end, five application programming interfaces (APIs) for every task were released together with the Jurassic-2 model, including paraphrase, summary, grammar check, long text classification by subject, and tools for recommending improvements in writing.

Particularly, each API is linked to a special version of the model optimized for every task in the bottom model Jurassic-2. For instance, summary models are trained on large amounts of summary data and use algorithms optimized for summarization. These working models are much smaller and more accurate than LLMs.

In line with AI21 Labs, the Jurassic-2 model has 178 billion parameters, which is greater than competing models resembling GPT-3, which has 175 billion, but has five times the vocabulary. This implies providing users with a way more accurate language solution.

AI21 Labs argued that in comparison with ChatGPT, whose current knowledge base is barely valid until the tip of 2021, Jurassic-2’s training dataset, updated by mid-2022, is more accurate than ChatGPT.

The Jurassic-2 language model is on the market in three sizes: Large, Grande and Jumbo, giving users more flexibility when it comes to cost and functionality depending on their needs. Depending on the model used, cost, latency and results vary. Other LLMs, including ChatGPT, offer only one-size-fits-all models with fixed cost and latency.

“The Jurassic-2 family of models represents a next-generation breakthrough and can enable developers and organizations to construct text-based applications faster, with peak performance, and at significantly less cost,” said Ori Gossen, CEO of AI21 Labs.

Chan Park, cpark@aitimes.com