AI21 Labs

AI21 Labs is a Tel Aviv-based artificial intelligence company specializing in natural language processing that develops systems able to understand and generate natural language. Its work spans two product lines: proprietary and open-weight foundation models for enterprise use — the Jurassic-1, Jurassic-2 and Jamba families, distributed via the AI21 Studio developer platform — and Wordtune, a consumer-facing AI reading and writing assistant that produces context-based rewrites, expansions and document summaries. The company transitioned its primary focus to enterprise applications in 2021, though Wordtune remained on the market as of August 2025.

The company's current positioning centers on Maestro, an optimization and orchestration framework for production AI agents launched in March 2025. AI21 describes Maestro as routing, compressing and decomposing agent requests in flight to reduce cost without code changes, mapping cost/accuracy/latency trade-offs, and attributing spend to specific teams, repositories and agents. Its stated system components include inference-time execution strategies, harness optimization, intelligent routing across an ensemble of models, and Jamba model training. AI21 also describes a neuro-symbolic architecture that combines large language models with external knowledge sources and discrete reasoning modules, and Task-Specific Models targeting narrow NLP tasks such as grounded question answering and summarization in order to limit hallucination.

The company maintains an active research output, publishing technical work on topics including hybrid Transformer-Mamba models, agentic evaluation at scale, RAG retrieval, vLLM debugging and scaling, online RL efficiency, and an AI Code of Conduct.

Founding story

AI21 Labs was founded in November 2017 in Tel Aviv, Israel by Amnon Shashua (chairman and co-founder), Yoav Shoham (co-CEO and co-founder) and Ori Goshen (co-CEO and co-founder); one source additionally lists Maykel Zayton among the founders. Shashua is a Hebrew University computer science professor who previously co-founded Mobileye — which held the record for the largest Israeli IPO in the US in 2014 and was acquired by Intel for $15.3 billion in 2017 — as well as OrCam, the OneZero digital bank (also an AI21 customer) and Mentee Robotics. Shoham is a Stanford professor emeritus of computer science and former Google principal scientist who previously founded TradingDynamics (acquired by Ariba), Timeful (acquired by Google) and Katango (acquired by Google). Goshen served in the IDF's 8200 unit, worked at the mobile communications startup Fring, and in 2009 founded the telecommunications analytics firm Crowdx, which served AT&T and Verizon before being acquired. Goshen and Shoham were introduced through Goshen's wife, who ran a programming bootcamp with Shoham; Shashua joined shortly after. At a time when much AI research focused on computer vision and deep learning, the founders bet on large language models, and the company emerged from stealth in October 2020 with Wordtune.

Business model

AI21 operates a software business combining B2B enterprise model licensing and deployment with a consumer application. Reported revenue channels include premium consumer subscriptions for Wordtune, API usage fees through AI21 Studio, and enterprise licensing. Models are offered for flexible deployment ranging from air-gapped and private enterprise installations to hybrid cloud, and are distributed through cloud marketplaces including Amazon Bedrock and Google Cloud.

Premium consumer subscriptions (Wordtune, a freemium editor), API usage fees, and enterprise licensing for corporate customers, per third-party profile data.

Traction

Wordtune is reported to have around 10 million users worldwide and to have delivered more than 782 million rewrite suggestions as of August 2025, supporting 10 languages. Headcount is reported at roughly 200-250 employees (251 per Contrary Research as of September 2025; 201-500 per Startup Intros). Total disclosed funding reached $336 million at a $1.4 billion valuation in November 2023, with one third-party source citing $626.5 million total funding and a Series D stage as of September 2025.

Latest developments

Recent published research and product messaging (dated into 2026 on AI21's own site) centers on agentic AI: orchestrated test-time compute for long-horizon tasks, MCP workspaces for state-modifying agents, scaling agentic evaluation across 200,000 SWE-bench runs, budget-aware Best-of-N execution for SWE agents, verifier-based approaches, and pairing open models with frontier models for patching. Commercially, the company presents Maestro as a cost-optimization framework for production AI agents, claiming most agent budgets contain recoverable waste.