AI21 Labs

What the company is

AI21 Labs is an Israeli artificial intelligence company founded in November 2017 in Tel Aviv by Yoav Shoham, Ori Goshen and Amnon Shashua^[1]. Ori Goshen is chief executive officer, and Yoav Shoham, a Stanford professor emeritus, is chairman^[2]. The company describes Goshen as a seasoned entrepreneur with more than 15 years of experience in technology and product leadership roles^[2].

Its stated mission is to "enable every company to own its specialized, frontier-level intelligence," and its site puts the same idea another way: "Intelligence should be an asset you grow, not an expense you meter"^[2]. In practice the company has worked on two fronts at once: a developer platform and family of large language models, and writing tools aimed at end users.

Writing tools

One of its first widely known products was Wordtune, an AI writing assistant launched on 27 October 2020^[1]. In 2021 Google named it one of its favourite Chrome extensions of that year^[1]. A follow-up product, Wordtune Spices, arrived on 17 January 2023^[1].

The Jurassic models and AI21 Studio

In August 2021 the company launched AI21 Studio, a developer platform for building on its models, together with the Jurassic-1 family^[1][3]. The largest model in that family, Jurassic-1 Jumbo, had 178 billion parameters^[1][3]. Parameters are the adjustable numerical weights a model learns during training; a larger count generally means more capacity, though it does not by itself decide how useful a model is. Jurassic-2 was released on 9 March 2023^[1].

Jamba: a hybrid architecture

On 28 March 2024 the company released Jamba, an open-weights model — meaning its weights are published for others to download and run — that combines Mamba state-space layers, Transformer layers and mixture of experts^[1][4]. The three ingredients do different jobs. Transformer attention lets every token compare itself with other tokens in the sequence, which is expressive but grows costly as sequences get longer. A state-space layer instead carries a compressed running state forward, which is cheaper over long inputs. Mixture of experts routes each token to only a subset of specialised sub-networks, so only part of the model is active for any given token. Stacking them is an attempt to get long-context efficiency without giving up the strengths of attention.

In August 2024, Jamba 1.5 scaled the hybrid design to 398 billion total parameters^[5], and the model was offered on Amazon Bedrock through a partnership with AWS^[1][5].

In March 2025 the company released Jamba 1.6 for private enterprise deployment and launched Maestro, a planning and orchestration system for AI agents^[1].

Funding and investors

The company raised a $9.5 million seed round in January 2019, a $25 million Series A led by Pitango First in November 2021, a $64 million Series B led by Ahren in July 2022, and a $155 million Series C in August 2023 that brought in new investors including Google and Nvidia^[1]. Investors listed by the company include Google, Nvidia, Intel Capital, Samsung Next, Coatue, Pitango, Ahren and Comcast Ventures^[2].

Where it sits

Across its releases AI21 Labs has covered three layers at once: writing tools for end users, a developer platform and model family for builders, and open-weights models plus private deployment for enterprises^[1]. Maestro, launched in March 2025, moves the emphasis further toward orchestrating agents rather than only supplying a model^[1]. The company's own framing — "Intelligence should be an asset you grow, not an expense you meter" — matches that direction: models that customers run and build on themselves^[2].