AI21 Labs

AI21 Labs is an Israeli artificial intelligence company headquartered in Tel Aviv, founded in 2017 by Stanford professor emeritus Yoav Shoham, entrepreneur Ori Goshen, and Mobileye founder Amnon Shashua.¹² It is one of the longest-running independent large language model developers outside the United States and China, and its history traces the industry's arc in miniature: a frontier-scale dense Transformer line (Jurassic, whose 2021 debut model was among the largest then available for paid public use), a widely used consumer writing product (Wordtune), a distinctive architectural bet (Jamba, the first production-scale attention/state-space hybrid), and finally a pivot to enterprise "AI systems" centered on the Maestro orchestration layer.³⁴⁵

Origins

AI21 Labs was founded in Tel Aviv in 2017, before the Transformer wave had produced GPT-2, by three figures with unusually deep academic and industrial credentials: Shoham, a professor emeritus at Stanford known for work in game theory and multi-agent systems and a former principal scientist at Google; Goshen, who had previously co-founded the network analytics company Crowdx; and Shashua, the founder of the driver-assistance company Mobileye.¹² The company's name gestures at its founding thesis, "AI for the 21st century," and its early publications argued that statistical language models would need to be combined with reasoning, tools, and structure rather than scaled in isolation, a stance Shoham maintained even as the company fielded some of the era's largest models.

The company raised venture funding in stages as the generative AI market inflated: a $64 million round in July 2022, then a Series C announced at $155 million in August 2023 at a reported $1.4 billion valuation, closed oversubscribed at $208 million that November with backers including Google, NVIDIA, Intel Capital, and Samsung Next, bringing total funding to roughly $336 million.²⁶

Model program

Jurassic

AI21's first model generation, Jurassic-1, was announced in August 2021. Its largest variant, Jurassic-1 Jumbo, was developer-reported at 178 billion parameters, marginally larger than OpenAI's GPT-3 and, at launch, among the largest dense models publicly available for paid use alongside contemporaries such as Megatron-Turing NLG.³ The model shipped with deliberate architectural departures from the GPT-3 recipe, including a shallower, wider layer configuration justified by the lab's own depth-to-width research and a roughly 256,000-item tokenizer vocabulary containing multi-word expressions. Jurassic-1 anchored AI21 Studio, the company's developer API, and positioned AI21 with Cohere and Aleph Alpha in the small club of non-US-giant labs selling frontier-scale text models. A 2022 research system, Jurassic-X, implemented the lab's "MRKL" proposal for routing queries from the language model to calculators, databases, and search, anticipating the later industry turn to tool use. Jurassic-2 followed in March 2023 with undisclosed parameter counts, multilingual coverage, instruction-tuned variants, and distribution through Amazon Bedrock alongside Amazon Titan.³

Jamba

In March 2024 the company retired the dense-Transformer line in favor of Jamba, the first production-scale model to interleave Transformer attention layers with Mamba state-space layers under a mixture-of-experts feed-forward design: 52 billion total parameters with about 12 billion active, a 256K context window, and open weights under Apache 2.0.⁴ Jamba made the hybrid architecture question concrete for the wider field; interleaved attention-SSM designs subsequently appeared across the industry, from NVIDIA's Nemotron-H line to MiniMax's efficient-attention models. Jamba 1.5 (August 2024) scaled the family to a 398-billion-parameter Large variant, with 1.6 and 1.7 refreshes through 2025 emphasizing grounded answers and self-hosted deployment, distributed via Hugging Face and the company's own API.⁴ In October 2025, Jamba Reasoning 3B pushed the recipe down-market: a small open reasoning model handling more than 250,000 tokens of context on laptop-class hardware, which VentureBeat framed as redefining what "small" means in LLMs.⁷

Products and the enterprise pivot

AI21's first product was consumer-facing: Wordtune, an AI writing assistant offering context-aware paraphrases and rewrites, launched on October 27, 2020, well before ChatGPT made such tools ubiquitous.² Over time the company's center of gravity shifted decisively toward enterprise infrastructure. In March 2025 it introduced Maestro, marketed as an "AI planning and orchestration system" for the enterprise: a layer that plans, executes, and validates multi-step tasks around third-party models as well as AI21's own, with the company reporting accuracy gains of up to 50 percent for paired models such as GPT-4o and Claude Sonnet 3.5 (developer-reported figures).⁵ Maestro became generally available for deployment in Amazon Virtual Private Cloud in December 2025.⁵ The pivot echoed the founders' original systems-over-models thesis: the company now describes itself as building "enterprise AI Systems and Foundation Models" rather than competing for the raw-capability frontier.¹

Stated mission and record

AI21's about page states its aim as "Trustworthy artificial intelligence that powers humanity towards superproductivity," and describes the company as "pioneering the development of enterprise AI Systems and Foundation Models."¹ The framing reflects a long-held house position, argued publicly by Shoham since the Jurassic-X and MRKL work, that language models alone are insufficient and must be embedded in systems that plan, verify, and use tools; as of July 2026 the company has published no separate safety framework or charter comparable to those cataloged at Stated missions of AI labs.¹⁵

Parts of the record track the stated emphasis on trustworthiness and systems: the Jamba family shipped as open weights under Apache 2.0 with an accompanying technical paper, Maestro is explicitly built around validation of outputs, and the company's tool-routing research predated the industry consensus on the problem.⁴⁵ Other parts cut the other way or remain unverifiable: the Jurassic models were proprietary and their successors' parameter counts undisclosed, headline claims such as Maestro's 50 percent accuracy gains and Jamba's benchmark parity are developer-reported, "superproductivity" is a slogan without measurable commitments attached, and the consumer product that most directly served everyday users, Wordtune, receded from the company's positioning as it pursued enterprise revenue. The mission's operative terms describe engineering reliability rather than safety governance, and the public evidence for the "trustworthy" label consists mostly of the company's own materials and benchmarks.

Position in the field

AI21 occupies an unusual position: too small to contest the capability frontier against OpenAI, Anthropic, or Google DeepMind, but too technically distinctive to be dismissed as a wrapper company. Its closest analogues are other national-champion independents such as Mistral AI in France, Cohere in Canada, and the Technology Innovation Institute in the UAE, each balancing open releases against enterprise sales. The lab's durable contribution has been architectural: Jamba demonstrated that attention-SSM hybrids could ship at production scale, and the hybrid template it validated now appears throughout the efficiency-focused end of the model landscape.⁴⁷ Its Maestro bet, that orchestration and validation layers rather than base models will capture enterprise value, remained commercially unproven but directionally aligned with the broader 2025-2026 agentic turn.⁵

See also

• Jurassic (models)
• Jamba
• Hybrid architecture (LLM)
• Stated missions of AI labs

References