AI21 Labs

From language research to specialized models

AI21 was founded in 2017 by Amnon Shashua, Yoav Shoham and Ori Goshen. It combined research into language models with applications for reading and writing. Its 2022 financing announcement identified Tel Aviv as its headquarters.

The current leadership page names Goshen as CEO and Shoham as chairman. Goshen previously co-founded network analytics company Crowdx. Shoham is a Stanford computer science professor emeritus and a former Google principal scientist.

Jurassic and Jamba use different designs

Jurassic-1 used decoder-only Transformers. The original technical paper documented 7.5-billion and 178-billion-parameter models with a 256K-item vocabulary.

Jamba arrived in March 2024 with a combination of attention, Mamba state-space layers and mixture-of-experts computation. The design addressed the memory and processing demands of long text.

Jamba2 followed in January 2026 with a dense 3B model and a Mini model using 12 billion active parameters. Both releases use Apache 2.0, unlike the custom license on the Jamba 1.5 to 1.7 weights.

A disclosed valuation belongs to a specific financing round

AI21 completed a $208 million Series C on 21 November 2023 at a stated $1.4 billion valuation. The completed round brought disclosed funding to $336 million. Intel Capital, Comcast Ventures and Ahren joined the extended round.

Practical guidance. That figure values the private company at the time of the transaction. It is not a current stock-market capitalization, an annual revenue figure or a public price for the Jamba model family.

Model developers and hosted models are different roles

AI21's Maestro updates include integrations with Mistral models and Together AI. Those integrations expand its service options; they do not make the external models AI21-trained checkpoints.

Practical guidance. The model list below separates AI21's named releases from orchestration products. Wordtune and Maestro are products, while Jurassic-X describes a modular system. A service or task name is not counted as an additional foundation-model release without a distinct model identity.

How the model list handles versions

Practical guidance. The list includes public Jamba checkpoints, documented historical Jurassic releases, the Jamba 1.5a research adaptation and the named Contextual Answers model. FP8 and GGUF exports appear as formats or aliases on their parent model pages. Random initialization repositories and synthetic development reward heads are excluded.

Practical guidance. Access labels distinguish licensed weights from hosted APIs. Older launch announcements do not establish that an endpoint still accepts new users. Where a current access route is not confirmed, the page states that limitation.