AI21 Labs

  Executive Briefing

AI21 Labs is an Israeli enterprise AI company founded in November 2017 in Tel Aviv, built on the conviction that language — not computer vision — was the true frontier of artificial intelligence. It is best known for producing the Jurassic family of large language models (2021–2023) and the Jamba series (2024–2025), the latter distinguished by a landmark hybrid Mamba state-space model and Transformer architecture that offered dramatically more memory-efficient long-context processing than contemporaneous alternatives. By May 2026, after failing to close a reported $300 million Series D, watching multiple acquisition talks fall through, and watching revenue stagnate at an estimated $50 million annually, the company executed a radical pivot: it discontinued standalone model sales, shuttered its Wordtune consumer product, cut more than 60% of its workforce, and bet the company's future on Maestro, a model-agnostic AI agent optimization and orchestration platform.¹

The company is led today by co-founders Prof. Yoav Shoham (Co-CEO) and Ori Goshen (Co-CEO), with Prof. Amnon Shashua serving as Chairman and Barak Lenz as CTO. Shoham is a professor emeritus at Stanford and former Google Principal Scientist whose prior ventures — TradingDynamics, Timeful, and Katango — were each acquired by large technology companies. Shashua is a Hebrew University of Jerusalem computer science professor and the co-founder of Mobileye, acquired by Intel in 2017 for $15.3 billion, making AI21 his second consequential technology venture. Goshen is a serial entrepreneur who previously co-founded CrowdX (a telecoms analytics company) and was part of the early team at Fring, a mobile VoIP startup that scaled to 40 million users.²

At its peak, AI21 was one of a handful of companies globally capable of training frontier-scale language models, commanded a $1.4 billion valuation, and counted Google and NVIDIA among its investors. The May 2026 restructuring — reducing headcount from roughly 180 to approximately 70 employees — reframed the company's ambition: rather than competing at the model frontier, AI21 now positions itself as infrastructure for enterprises that want reliable, production-grade AI agent systems regardless of which underlying model they use. Whether this pivot constitutes a durable strategy or a managed wind-down is the defining question about AI21 as of mid-2026.

  At a Glance

  Origins & Founding

AI21 Labs was incorporated in November 2017 in Tel Aviv by four co-founders: Prof. Yoav Shoham, Ori Goshen, Prof. Amnon Shashua, and Maykel Zayton. The name is a deliberate play on "AI for the 21st century." The founders made a contrarian choice at formation: rather than competing in computer vision — the then-dominant battleground for AI — they staked the company's identity on natural language processing, arguing that language most directly reflected human cognition and would prove the more consequential and underexplored frontier.²

The founding thesis was explicitly neurosymbolic: the team believed that pure deep-learning pattern matching was insufficient for reliable, high-stakes AI systems, and that the path to trustworthy AI required integrating deep learning with symbolic reasoning — structured logic, planning, and explicit knowledge representation. This philosophical stance would shape the company's product roadmap for nearly a decade, ultimately manifesting in Maestro's plan-based orchestration approach. The company remained in stealth for three years, raising a $9.5 million seed round in January 2019, before publicly emerging in October 2020 with the launch of Wordtune, an AI writing assistant.⁴

Barak Lenz, who had served alongside Goshen in the Israeli Defense Forces' Intelligence Corps, joined as co-founder and CTO from the outset. The founders' combined credentials — Stanford-and-Google pedigree (Shoham), a $15.3B technology exit (Shashua), and entrepreneurial execution experience (Goshen, Zayton, Lenz) — gave the team unusually deep access to both academic AI talent and enterprise customers from day one.²

  History & Timeline

    2017–2020: Stealth and the Wordtune debut

The company operated in stealth for its first three years, conducting foundational research into large-scale language modeling and assembling a team of NLP researchers. A $9.5 million seed round in January 2019 funded this period. The company publicly emerged in October 2020 with Wordtune, a browser-based AI writing assistant that could rephrase, expand, or shorten sentences while preserving the author's meaning. Wordtune quickly attracted millions of users as one of the earliest polished consumer demonstrations of what large language models could do for everyday writing tasks, ultimately reaching a reported 10 million peak users.⁴

    2021–2022: Jurassic models and the API platform

In August 2021, AI21 launched two foundational pieces of its enterprise strategy simultaneously: the AI21 Studio developer API platform and the Jurassic-1 model family. J1-Jumbo, at 178 billion parameters, was among the largest language models publicly available at the time and featured an unusually large 250,000-token vocabulary — compared to GPT-3's 50,000 — a design choice intended to improve tokenization quality for diverse text and reduce per-word token costs for enterprise customers.⁵ The smaller J1-Large at 7 billion parameters targeted cost-sensitive applications.

The Jurassic-1 launch attracted significant enterprise interest and enabled the company to raise a $25 million Series A led by Pitango First (November 2021) and a separate $20 million from Walden Catalyst at a reported $400 million valuation, followed by a $64 million Series B led by Ahren Innovation Capital in July 2022 with participation from Walden Catalyst, Pitango, TPY Capital, and investor Mark Leslie.³ The company's revenue model took shape: usage-based API pricing for enterprise customers in finance, healthcare, legal, and technology sectors who required reliable, high-volume text processing.

    2023: Jurassic-2, the unicorn round, and Wordtune Spices

March 2023 saw the launch of Jurassic-2, offering multilingual support across seven languages, 30% faster response times, and improved instruction-following — the company's direct answer to the competitive pressure unleashed by ChatGPT's November 2022 debut. The same month, Wordtune Spices was released as an add-on that could generate illustrative examples, counterarguments, and statistics directly within the writing assistant, extending the product from rewriting to generative assistance.⁴

The company's most consequential capital event arrived in August 2023, when AI21 announced a $155 million Series C at a $1.4 billion valuation — attaining unicorn status — with investors including Google, NVIDIA, Samsung Next, SCB10X, b2venture, Walden Catalyst, and Pitango.⁶ By November 2023, the round was extended and oversubscribed, closing at $208 million total with the addition of Intel Capital and Comcast Ventures, bringing confirmed total funding to approximately $336 million.⁷ This valuation and funding level would prove to be AI21's high-water mark.

    2024: The Jamba architectural bet

March 29, 2024 marked a significant technical departure: AI21 released Jamba, an open-weight hybrid model combining the Mamba state-space model (SSM) architecture with the Transformer architecture and a Mixture of Experts (MoE) routing mechanism.⁸ Jamba supported a 256,000-token context window — four times the context of most contemporaneous commercial models — while being reported as up to 10 times more memory-efficient than Llama 3 and Mistral for equivalent context lengths. It was the first major publicly available model to successfully combine SSM and Transformer paradigms in a production architecture, attracting substantial research attention.

August 22, 2024 brought Jamba 1.5 Mini (52 billion total parameters, 12 billion active) and Jamba 1.5 Large (398 billion total parameters, 94 billion active), with the Large variant reported as up to 2.5 times faster at inference than comparably sized dense models.⁸ A September 2024 partnership with AWS made the Jamba 1.5 family available on Amazon Bedrock, and the models were also listed on Microsoft Azure AI. Despite strong architectural novelty, Jamba's real-world benchmark performance placed it roughly in the middle tier of available models: ranked approximately 107–141 on the LM Arena leaderboard as of 2025.

    2025–2026: Failed fundraise, pivot to Maestro, and restructuring

Early 2025 brought reports of a $300 million Series D being negotiated with Google and NVIDIA; this round was widely reported as complete by several technology outlets in May 2025 but was never officially closed or formally announced by AI21.⁹ Calcalist (Ctech), the Israeli technology publication that has most closely tracked the company, reported that the Series D did not formally close and that the company's last confirmed capital raise remains the November 2023 Series C extension.¹ Uncertainty about this round is a material consideration: confirmed total funding stands at approximately $336 million, while the figure would reach approximately $636 million if the reported Series D is counted.

March 2025 saw the release of Jamba 1.6 for private enterprise deployment and the launch of Maestro, an AI planning and orchestration system designed as a model-agnostic layer capable of routing tasks across OpenAI, Anthropic, and AI21 models. In April 2025, Wordtune was quietly discontinued after more than four years as the company's consumer flagship. Reports in May 2025 described acquisition talks with NVIDIA at an estimated valuation of $2–3 billion; those talks ultimately did not result in a deal. Jamba Large 1.7 was released on HuggingFace in July 2025, improving context grounding and steerability.¹⁰

By April 2026, Nebius Group — the European cloud spin-off from Yandex — entered acquisition talks for AI21 after the NVIDIA deal fell through. On May 18, 2026, AI21 announced that Nebius acquisition talks had been terminated and that the company would instead sign commercial partnerships with both Nebius and Wix, with Nebius paying a reported "tens of millions of dollars" to license the Maestro platform for AI agent development.¹ Simultaneously, the company announced a workforce reduction of more than 60%, cutting headcount from approximately 180 to approximately 70 employees — predominantly developers — and formally pivoted its entire product strategy to Maestro, discontinuing standalone model sales. Ctech described the situation as "AI21's fall from unicorn to survival mode."¹

  Mission, Philosophy & Research Agenda

AI21's founding mission was framed as "AI for the 21st century" — a company that would advance artificial intelligence in the broadest humanistic sense, not merely optimize metrics. Its stated current mission is "Trustworthy artificial intelligence that powers humanity towards superproductivity." The emphasis on trustworthy is not incidental: the company's foundational philosophy holds that most AI failures in high-stakes enterprise environments stem not from insufficient raw capability but from insufficient reliability — models hallucinate, deviate from instructions, produce variable outputs across runs, and fail silently.

The neurosymbolic thesis that animated the company's founding found its most mature expression in the Maestro architecture. Co-CEO Shoham has articulated the platform's philosophy as: "Stop treating LLMs as standalone products and start building complete systems around them — systems that account for uncertainty, monitor outputs, manage costs, and layer in guardrails for safety." Maestro generates structured, verifiable plans for AI agent tasks before executing them, with human-reviewable intermediate representations — a direct application of the symbolic-reasoning strain of the founding thesis to agentic AI systems.

Research themes across AI21's history have consistently prioritized: (1) long-context processing and efficiency — from the 250K-token vocabulary of Jurassic-1 to Jamba's 256K-token window; (2) hybrid architectures that combine the strengths of transformers (recall, in-context reasoning) with state-space models (linear scaling with sequence length); (3) accuracy and reliability over raw benchmark ranking — the company consistently positions its models for enterprise SLAs rather than leaderboard position; and (4) explainability and auditability appropriate for regulated industries such as finance, healthcare, and defense. The advisory board — including Dan Jurafsky (Stanford), Christopher Ré (Stanford), and Sebastian Thrun (Stanford) — reflects a sustained engagement with academic NLP and ML research even as the company pursued commercial objectives.

  Research & Publications

AI21's most significant technical contributions have been in model architecture and long-context efficiency. The company has published openly on its core innovations and released several models as open weights.

• Jurassic-1 Technical Evaluation (2021, authors including Opher Lieber) — introduced the 178-billion-parameter autoregressive transformer with a 250,000-token vocabulary and provided zero-shot and few-shot comparisons against GPT-3, establishing the model as a peer to the then-frontier.⁵
• Jamba: A Hybrid Transformer-Mamba Language Model (March 2024) — landmark open-weight architecture combining Mamba SSM blocks, standard Transformer attention blocks, and Mixture of Experts routing. One of the first production-scale demonstrations of the SSM-Transformer hybrid, it showed a 256K-token context window was achievable with dramatically lower memory footprint than pure transformer models.⁸
• Jamba 1.5 (August 2024) — updated architecture achieving a reported 20-percentage-point improvement on the MATH reasoning benchmark (62% → 82%) in a subsequent November 2024 update, demonstrating that the hybrid architecture could absorb post-training improvements in reasoning as effectively as transformer-only models.
• Maestro (March 2025) — introduced a structured plan-generation and verification approach to AI agent orchestration, claiming accuracy improvements from 85% to 91.9% on GPT-4o tasks and from 88% to 95.2% on Claude Sonnet 3.5 tasks when Maestro's planning layer was applied.

AI21's open-weight releases — particularly the Jamba family — have contributed meaningfully to the research community's understanding of hybrid SSM-Transformer architectures, even as the company's frontier model ambitions were ultimately constrained by capital.

  Models & Products

AI21's product history divides cleanly into three eras: the Wordtune consumer era, the Jurassic/Jamba foundation model era, and the current Maestro agent platform era.

Jurassic family (2021–2023):

• Jurassic-1 (Aug 2021) — J1-Jumbo (178B parameters) and J1-Large (7B); 250,000-token vocabulary; trained on 300 billion tokens. Released via the AI21 Studio developer API.
• Jurassic-2 (Mar 2023) — next-generation model with multilingual support across 7 languages, 30% faster response times, and improved instruction-following.

Jamba family (2024–2025):

• Jamba (Mar 2024) — open-weight hybrid Mamba SSM-Transformer MoE model; 256K token context window; first major public SSM-Transformer hybrid at production scale.
• Jamba 1.5 Mini (Aug 2024) — 52B total parameters, 12B active; efficiency-focused variant for cost-sensitive deployments.
• Jamba 1.5 Large (Aug 2024) — 398B total parameters, 94B active; reported up to 2.5x faster inference than comparable dense models.
• Jamba 1.6 (Mar 2025) — enterprise private deployment; outperformed Mistral Large 2 and Llama 3.3 on Arena Hard, RAG, and long-context QA benchmarks per AI21's internal evaluations.
• Jamba Large 1.7 (Jul 2025) — improved context grounding and steerability; available on HuggingFace.¹⁰

Consumer and developer products:

• Wordtune (Oct 2020 – Apr 2025) — AI writing assistant and paraphrasing tool; reached a reported 10 million peak users; discontinued April 2025.
• AI21 Studio (Aug 2021 – May 2026) — developer API platform for deploying AI21 foundation models; now deprecated as standalone model sales are discontinued.

Current product:

• Maestro (Mar 2025 – present) — model-agnostic AI planning and orchestration platform; works with OpenAI, Anthropic, and AI21 models; generates structured, verifiable plans before executing agentic tasks; sole active product line as of May 2026.

  People

AI21 has always been unusually founder-led, with all four co-founders active in the company across nearly a decade. The leadership team combines rare depth: two professor co-founders with active research credentials, two operator co-founders with prior exits, and a CTO who has been in role since the company's first day.

Co-founders and current leadership:

• Prof. Amnon Shashua — Co-founder and Chairman. Professor of computer science at the Hebrew University of Jerusalem with more than 160 publications and 94 patents. Previously co-founded Mobileye (acquired by Intel for $15.3 billion in 2017), OrCam (assistive technology), and OneZero (digital bank). Widely regarded as one of Israel's most accomplished technology entrepreneurs.²
• Prof. Yoav Shoham — Co-founder and Co-CEO. Professor emeritus at Stanford University and former Google Principal Scientist, with core expertise in AI, game theory, and multi-agent systems. Previously founded TradingDynamics, Timeful, and Katango — each acquired by major technology companies. His multi-agent systems background directly shaped AI21's agent orchestration strategy.²
• Ori Goshen — Co-founder and Co-CEO. Serial entrepreneur who co-founded CrowdX (telecoms analytics, acquired) and was part of the early team at Fring (mobile VoIP, 40 million users). Grew up in Kfar Saba, Israel; served in the IDF Intelligence Corps alongside Barak Lenz.²
• Barak Lenz — Co-founder and CTO. Has served as CTO since AI21's founding in 2017. IDF Intelligence Corps veteran alongside Goshen; the technical anchor of the company across its full history.
• Maykel Zayton — Co-founder. Listed as a co-founder in corporate filings and some sources but largely absent from official company narratives; exact current role is not publicly confirmed.²

Other key roles:

• Nellie Stockhammer — CFO
• Barak Peleg — VP Technology & Architecture
• Hila Yakobovich — Director of HR

Advisory board includes Dan Jurafsky (Stanford, computational linguistics), Christopher Ré (Stanford, ML systems), and Sebastian Thrun (Stanford, robotics and AI), as well as academics from Hebrew University, the Technion, the University of British Columbia, and the University of Pennsylvania — reflecting the co-founders' deep roots in academic NLP.

The May 2026 restructuring reduced overall headcount from approximately 180 to approximately 70 employees, with the departures reportedly concentrated among developers. Some reports suggest that a portion of departing engineers transitioned to Nebius Group as part of the commercial partnership, though exact numbers are not confirmed.

  Funding, Ownership & Business

AI21 is an independent company with no parent organization. Its investor base is notable for including two of the world's most strategically significant AI infrastructure companies — Google and NVIDIA — alongside a cluster of deep-technology venture funds.

Confirmed funding history:

Confirmed total raised: approximately $336 million. Last confirmed valuation: $1.4 billion (November 2023).³⁶⁷

A $300 million Series D led by Google and NVIDIA was widely reported in technology media in May 2025. However, this round was never officially announced or formally closed by AI21; Calcalist reported in 2026 that the negotiations did not result in a completed raise.¹⁹ Any figures citing $636 million in total funding reflect this unconfirmed round and should be treated with caution.

Business model: API usage-based pricing for enterprise customers. Pricing for the Jamba family was set at $0.20 per million input tokens and $0.40 per million output tokens for Jamba Mini, and $2.00 per million input tokens and $8.00 per million output tokens for Jamba Large — competitive with, but not undercutting, Anthropic and OpenAI equivalent tiers. Annual revenue is estimated at approximately $50 million by Calcalist, though other sources cite a range of $57–112 million; these figures are unverified and stagnant relative to the broader generative AI market boom.¹ The company's last confirmed revenue state was described as significantly below peers: Mistral's revenue is estimated at approximately $400 million and Cohere's at approximately $200 million for comparison.

  Partnerships & Ecosystem

AI21's partnership strategy pursued broad cloud distribution for the Jamba family and is now centered on the Maestro platform's first major commercial contracts.

Cloud and infrastructure partners:

• AWS / Amazon Bedrock (Sep 2024) — Jamba 1.5 Mini and Jamba 1.5 Large deployed on Amazon Bedrock, enabling enterprise customers to access the models through AWS's managed infrastructure and security environment.
• Microsoft Azure — Jamba 1.5 Mini listed on the Azure AI model catalog.
• Google Cloud — listed as a cloud deployment partner; Google is also a Series C investor.
• NVIDIA — both a Series C investor and infrastructure partner; reported acquisition talks in May 2025 did not result in a deal.

Commercial partnerships (2026):

• Nebius Group (May 2026) — Former acquisition target; following the termination of acquisition discussions, Nebius signed a commercial partnership to license the Maestro platform for AI agent development, with reported payment of "tens of millions of dollars." Some departing AI21 developers may have joined Nebius as part of the transition.¹
• Wix (May 2026) — Partnership to build AI-powered services on the Maestro platform; Wix is an established Israeli technology company well-positioned as an early Maestro enterprise customer.

Integration partners: Snowflake (listed integration), plus broader ecosystem integrations through the AI21 Studio API.

Notable enterprise customers cited: Capgemini, Lightricks, Ubisoft, Latitude, OneZero, Rapyd, Fnac (reported 26% quality improvement and 40% latency decrease using Jamba 1.6 for inventory classification), Auchan, Easyway, and Glassdoor. The company's customer base has been concentrated in European and Israeli enterprises in finance, healthcare, technology, defense, and manufacturing — verticals with high reliability requirements and, in many cases, data-sovereignty needs met by on-premises or private-cloud deployments of open-weight Jamba models.

  Notable Events & Controversies

AI21's post-2024 history is primarily a story of external deal failures and internal restructuring rather than research controversy. Each development is significant for understanding the company's current position.

• Failed Series D (early–mid 2025) — Negotiations with Google and NVIDIA for a reported $300 million Series D round were widely reported as concluded but were never formally announced by AI21. The gap between media reports (treating the round as closed) and AI21's actual announcements (none) created lasting ambiguity about the company's financial position.⁹ This failure to close a large up-round appears to have been a proximate trigger for the events that followed.
• Wordtune discontinuation (April 2025) — After more than four years as AI21's public-facing consumer product and the vehicle through which millions of users first encountered the company, Wordtune was quietly shut down. The product never generated material revenue relative to the company's cost structure, and the shutdown signaled a full commitment to the enterprise B2B model.⁴
• NVIDIA acquisition talks (May 2025) — Reports from Israeli technology media described acquisition discussions between AI21 and NVIDIA at an estimated valuation of $2–3 billion. These talks did not result in a transaction; neither party confirmed the reported valuation.¹
• Nebius acquisition talks and collapse (2026) — Following the NVIDIA deal's failure, Nebius Group entered acquisition discussions with AI21 in April 2026. These talks also ended without a transaction — reportedly due to disagreements over terms and the strategic fit — and were formally terminated in May 2026, triggering the restructuring announcement.¹
• May 2026 workforce reduction — On 18 May 2026, AI21 announced it was cutting more than 60% of its workforce, reducing from approximately 180 to approximately 70 employees. The company simultaneously announced the pivot to Maestro as its sole product, the discontinuation of standalone model sales, and the commercial partnerships with Nebius and Wix. Calcalist described the situation as "AI21's fall from unicorn to survival mode."¹

  Competitive Position

AI21 occupied a specific but contested niche in the enterprise LLM market: open-weight models with superior long-context efficiency, targeting regulated industries with data-sovereignty and reliability requirements that made closed-API models from OpenAI or Anthropic problematic. Its principal technical differentiator was the Jamba hybrid SSM-Transformer architecture, which achieved longer effective context windows at meaningfully lower memory and compute cost than pure transformer architectures.

The competitive disadvantages were structural. AI21's confirmed total funding of approximately $336 million is a fraction of OpenAI's $40+ billion, Anthropic's $10+ billion, or even Mistral's roughly $1 billion in capital raised. Annual revenue of approximately $50 million compares unfavorably to Mistral's estimated $400 million and Cohere's estimated $200 million. Jamba's leaderboard rankings — approximately 107–141 on LM Arena as of 2025 — placed it well outside the set of models enterprise buyers default to. Estimated market share in US businesses is reported at less than 0.1%, compared to 32.4% for OpenAI.¹

The May 2026 pivot to Maestro represents an acknowledgment that AI21 cannot compete at the model frontier with its current resource base. The repositioning — from model vendor to model-agnostic agent infrastructure — targets the gap between raw model capability and reliable production deployment, a problem that exists regardless of which frontier model an enterprise has licensed.

  Outlook & Roadmap

As of June 2026, AI21 Labs is a company of approximately 70 employees with a single active product — Maestro — and its first major commercial revenue contracts in place with Nebius and Wix. The company's near-term roadmap, as announced in March 2025, includes a planned Python-based domain-specific language for describing and composing agentic plans — a developer-facing interface intended to make Maestro's plan-generation and verification capabilities programmable by enterprise engineering teams.¹

The strategic bet underlying the pivot is that the enterprise AI market will bifurcate: large organizations will use multiple frontier models from multiple providers (OpenAI, Anthropic, Google, and others) and will need a reliable, auditable, model-agnostic orchestration layer to manage them in production. Maestro is positioned to fill this role, with the claimed accuracy improvements — 85% to 91.9% on GPT-4o tasks, 88% to 95.2% on Claude Sonnet 3.5 tasks — as the core performance narrative. The commercial partnership with Nebius provides both early revenue and a reference customer for the platform's use in AI agent development infrastructure.

Several risks define the near-term outlook. First, AI21 enters the agent orchestration market against well-resourced competitors including LangChain, LlamaIndex, and the native agentic offerings from OpenAI, Anthropic, and Google. Second, the company's technical team — its core asset — has just undergone a 60% reduction, and retention of the remaining ~70 employees is an open question. Third, the commercial deals with Nebius and Wix, while significant, must scale into a durable revenue base. Acquisition by a larger technology or cloud company remains a plausible outcome; Google, NVIDIA, Nebius, and Israeli defense and technology conglomerates have all been cited in various deal contexts.

---

  References

1. AI21 Labs restructuring and pivot, May 2026 — Calcalist / Ctech
2. Founder story: Ori Goshen — Pitango
3. AI21 Labs — Wikipedia
4. AI21 Labs company profile — Contrary Research
5. Jurassic-1 language models — AI21 Labs
6. AI21 Labs Series C announcement — AI21 Blog
7. AI21 Labs completes $208M oversubscribed Series C — AI21 Blog
8. Announcing Jamba — AI21 Blog
9. AI21 Labs raises $300M with Google, Nvidia — SiliconAngle (reported, not confirmed by AI21)
10. AI21-Jamba-Large-1.7 — HuggingFace
11. AI21 Labs additional Calcalist coverage
12. AI21 Labs additional Calcalist coverage

  References

1. AI21 Labs restructuring, pivot to Maestro, and workforce reduction — Calcalist, May 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

2. Founder backgrounds and founding story — Pitango founder profile and Wikipedia: AI21 Labs. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. Series C and total funding history — AI21 Blog: Series C and AI21 Blog: Series C extension. ↩ ↩² ↩³

4. Company history and Wordtune — Contrary Research: AI21 Labs. ↩ ↩² ↩³ ↩⁴

5. Jurassic-1 technical details — Stanford FMTI AI21 Report. ↩ ↩²

6. Series C at $1.4B valuation — AI21 Blog: Series C announcement. ↩ ↩²

7. Series C extended to $208M — AI21 Blog. ↩ ↩²

8. Jamba architecture — AI21 Blog: Announcing Jamba and HuggingFace: Jamba Large 1.7. ↩ ↩² ↩³

9. Reported (unconfirmed) Series D — SiliconAngle, May 2025. ↩ ↩² ↩³

10. Jamba Large 1.7 release — HuggingFace. ↩ ↩²