# AI21 Labs

> Source: https://aiwiki.ai/wiki/ai21_labs
> Updated: 2026-10-28
> Fact-checked: 2026-10-27
> Categories: AI Companies, Large Language Models, Natural Language Processing
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "AI21 Labs." aiwiki.ai, 28 Oct 2026. https://aiwiki.ai/wiki/ai21_labs
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**AI21 Labs** is an Israeli [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company, founded in 2017 by [Yoav Shoham](https://aiwiki.ai/wiki/yoav_shoham), Ori Goshen, and [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua), that develops [large language models](https://aiwiki.ai/wiki/large_language_model) (LLMs) and AI systems for enterprise use. Headquartered in Tel Aviv, it is best known for the [Jamba](https://aiwiki.ai/wiki/jamba) model family, which AI21 introduced in March 2024 as the first production-grade [language model](https://aiwiki.ai/wiki/language_model) built on a hybrid [Mamba](https://aiwiki.ai/wiki/mamba) [state space model](https://aiwiki.ai/wiki/state_space_model) (SSM) and [Transformer](https://aiwiki.ai/wiki/transformers) architecture (the name Jamba stands for "Joint Attention and Mamba").[3][4] Its earlier Jurassic series and the [Wordtune](https://aiwiki.ai/wiki/wordtune) consumer writing assistant established the company before its architectural pivot.[1][11] AI21 Labs last closed funding in November 2023, an oversubscribed $208 million Series C that brought total funding to $336 million at a $1.4 billion valuation; a $300 million round reported in 2025 was never completed.[9][17] Press reports of acquisition talks with [NVIDIA](https://aiwiki.ai/wiki/nvidia) in late 2025, which AI21 denied, and with [Nebius](https://aiwiki.ai/wiki/nebius) in 2026 did not lead to a sale.[22][25][27] In May 2026 the company cut its workforce from about 180 to about 70 and said it would stop selling standalone models, and by October 2026 it was marketing post-training of specialized models for enterprises as a "lab-as-a-service".[27][28][33]

<figure class="w-56">
<img src="/images/site-2026-10-28-ai21_labs-logo-28246bce077d-tile.png" alt="AI21 Labs logo" width="214" height="93" />
<figcaption>AI21 Labs logo, from the official website <a href="https://ai21.com/">ai21.com</a>. Trademark of its owner.</figcaption>
</figure>

## What does AI21 Labs do?

AI21 Labs has built [foundation models](https://aiwiki.ai/wiki/foundation_models) and [enterprise AI](https://aiwiki.ai/wiki/enterprise_ai) infrastructure with a focus on accuracy, grounding, and long-context reasoning. Until 2026 its portfolio spanned three layers: foundation models (the Jurassic and Jamba families), an AI planning and orchestration system (Maestro), and developer and consumer products (AI21 Studio, Task-Specific APIs, and Wordtune). The company positioned itself against general-purpose providers like [OpenAI](https://aiwiki.ai/wiki/openai) and [Anthropic](https://aiwiki.ai/wiki/anthropic) by emphasizing reliability and deployment flexibility for enterprise workflows.

In May 2026 AI21 said it would concentrate on optimizing [AI agents](https://aiwiki.ai/wiki/ai_agents) around Maestro and discontinue the sale of standalone models.[27][28] By October 2026 its website listed two products: a post-training service, in which AI21 researchers train specialized models on a customer's own data, and AI21 Gateway, a cost-reduction endpoint for agent traffic that was in beta.[33][35]

## History

### Founding and early years (2017-2020)

AI21 Labs was founded in 2017 in Tel Aviv, Israel, by three co-founders with backgrounds in artificial intelligence research and technology entrepreneurship.[8]

**Yoav Shoham** is a professor emeritus of computer science at [Stanford University](https://aiwiki.ai/wiki/stanford_university) and a former Principal Scientist at [Google](https://aiwiki.ai/wiki/google). He has made notable contributions to AI, game theory, and [multi-agent system](https://aiwiki.ai/wiki/multi_agent_system)s over a career spanning several decades.[32] He previously sold two startups to Google, the time-management app Timeful and the social-network organizer Katango.[8] As of October 2026, AI21's website listed him as the company's chairman.[32]

**Ori Goshen** is a technology entrepreneur with over 15 years of experience in product leadership roles. Before co-founding AI21 Labs, Goshen co-founded Crowdx, a network analytics company, and led the development of VoIP products.[32] As of October 2026, the company's website listed him as CEO.[32]

**Amnon Shashua** is a professor of computer science at the Hebrew University of Jerusalem and a prominent Israeli entrepreneur. He co-founded [Mobileye](https://aiwiki.ai/wiki/mobileye), the [autonomous driving](https://aiwiki.ai/wiki/autonomous_driving) technology company that [Intel](https://aiwiki.ai/wiki/intel) acquired in 2017 for approximately $15.3 billion. Shashua also co-founded OrCam, a company focused on assistive technology for the visually impaired, and the digital bank One Zero. He was AI21's chairman until 2026: Calcalist reported in July 2026 that he intended to leave the company and its board, and Globes reported in August that he had announced his retirement as chairman.[30][31]

The company's name, AI21, refers to "artificial intelligence for the 21st century."[13] The founders started with a broad vision of building AI systems that could genuinely understand and generate natural language, rather than targeting a specific product from the outset. In 2019, AI21 Labs raised $9.5 million in a seed funding round.[43] Through its early years, the company operated in stealth mode while developing its core [natural language processing](https://aiwiki.ai/wiki/natural_language_processing) technology.

### Wordtune launch and coming out of stealth (2020)

On October 27, 2020, AI21 Labs emerged from stealth with the launch of Wordtune, its first consumer product.[11] Wordtune is an AI-powered writing assistant that can understand the context and meaning of text and suggest paraphrases, rewrites, and alternative phrasings. The product launched as a Chrome browser extension and quickly gained traction among individual users and professionals.[11] The following month, AI21 Labs completed a $25 million Series A round led by Pitango First, with participation from TPY Capital; Calcalist reported that the company then had 40 employees.[43]

### Scaling language models (2021-2023)

In 2021, AI21 Labs began to expand rapidly. In August 2021 the company launched AI21 Studio and its Jurassic-1 family of language models, entering the market alongside OpenAI and other foundation model developers.[1] In November 2021, the deep-tech venture firm Walden Catalyst invested $20 million in the company, an investment Calcalist reported at a $400 million valuation.[44]

In July 2022, AI21 Labs raised $64 million in a Series B round led by Ahren Innovation Capital, bringing its total funding to $118 million at a valuation of $664 million.[10] The funds were directed toward research and development, as well as expanding the company's sales and marketing teams.[10]

In March 2023, the company released Jurassic-2, a significant upgrade to its model family.[2] By August 2023, AI21 Labs had raised $155 million in the first tranche of its Series C round, reaching a valuation of $1.4 billion and achieving unicorn status. Investors in this round included Walden Catalyst, Pitango, SCB10X, b2venture, Samsung Next, and co-founder Amnon Shashua, with Google and NVIDIA also participating.[8]

In November 2023, AI21 Labs extended its Series C with an additional $53 million, bringing the total Series C to $208 million. New investors Intel Capital and Comcast Ventures joined the round, and the company's total funding reached $336 million.[9]

In September 2023, Shoham told VentureBeat that the company had grown from 50 to 200 employees and planned to add another 100 the following year.[13] Calcalist later reported that headcount had been about 260 in 2023 and that the plan to hire 100 more people in 2024 never materialized.[17]

### The Jamba era (2024-2026)

In March 2024, AI21 Labs released Jamba, a model that marked a significant departure from the pure Transformer architecture that had dominated the LLM landscape. Jamba introduced a hybrid architecture combining Mamba state space model (SSM) layers with Transformer attention layers. According to AI21, this made Jamba the first hybrid model of its kind to reach production-grade scale.[3]

In August 2024, the company released the Jamba 1.5 model family, scaling the hybrid architecture further with Jamba 1.5 Large (398B total parameters) and Jamba 1.5 Mini (52B total parameters).[5]

In March 2025, AI21 Labs released Jamba 1.6, which it marketed for private deployment on premises or in a virtual private cloud,[37] and introduced Maestro, an AI planning and orchestration system.[19] In April 2025 the company halted development of Wordtune, according to Calcalist.[12][29]

The Jamba 1.7 models followed; AI21's documentation dates the API snapshots of both Jamba 1.7 Large and Jamba 1.7 Mini to July 2025. They were incremental updates that improved grounding and instruction-following while retaining the 256K [context window](https://aiwiki.ai/wiki/context_window) and the nine supported languages, including Arabic and Hebrew, that the family had offered since Jamba 1.5.[16][5] Maestro became available to enterprise customers in July 2025,[39] and in October 2025 the company released [Jamba Reasoning 3B](https://aiwiki.ai/wiki/jamba_reasoning_3b), a small open reasoning model.[38] Late in 2025 AI21 announced a partnership with [Together AI](https://aiwiki.ai/wiki/together_ai) to connect Maestro to Together's hosted open-source models,[41] general availability of Maestro for deployment in Amazon virtual private clouds,[40] and a partnership with the French retailer Fnac Darty to use Maestro for after-sales service.[42]

In January 2026, AI21 Labs released [Jamba 2](https://aiwiki.ai/wiki/jamba_2), a pair of compact models (3B and Mini) focused on enterprise reliability, grounding, and instruction-following, released under the Apache 2.0 license.[6]

### What happened with the reported Series D and NVIDIA talks (2025-2026)?

In May 2025, Globes, citing Business Insider, reported that AI21 Labs was raising $300 million, a round that would have brought its total funding to $636 million.[45] Through much of 2025 the company was widely reported to have raised, and in some accounts to have closed, a $300 million Series D backed by Google and NVIDIA. According to later reporting by Calcalist, that round was never completed, never formally announced, and never reflected in the company's capital structure; AI21's last actual fundraising remained the November 2023 Series C extension to $208 million at a $1.4 billion valuation.[17] Globes reported in January 2026 that AI21, NVIDIA and Google had discussed and agreed on a $300 million investment that was not completed, leaving the company with no more than $336 million raised since its founding.[22] According to Calcalist, the company had said at the time that discussions which began as investment talks had evolved into acquisition talks, causing delays.[29] (A Globes article of December 30, 2025 had instead described AI21 as having raised almost $700 million, including a round in May 2025 at a $1.4 billion valuation.[21])

On December 30, 2025, Calcalist reported that NVIDIA was in advanced talks to acquire AI21 Labs for $2 billion to $3 billion. The reported deal was described as primarily an acqui-hire, with NVIDIA seeking to secure AI21's team of roughly 200 employees, most of whom hold advanced degrees and possess specialized expertise in AI research. If completed, the acquisition would have represented NVIDIA's fourth significant purchase in Israel and its second-largest after the $7 billion [Mellanox](https://aiwiki.ai/wiki/mellanox) acquisition in 2020.[12] Calcalist also cited estimates that AI21's annual revenue was about $50 million, and other outlets, including Verdict and SiliconANGLE, repeated the report.[12][23][24] The same day, Globes reported that AI21 had begun looking for a buyer and was seeking more than $2 billion, and quoted an internal email in which Goshen told employees that the company was "in talks with a number of potential entities, including Nvidia" but that this did "not indicate a specific deal or any immediate change."[21]

In an interview with Bloomberg in early January 2026, Shashua acknowledged the discussions while playing them down, saying: "There are talks with Nvidia, there are talks with others, but nothing even close to talking about it in the press."[18] On January 12, 2026, AI21 issued a statement saying that "reports of negotiations for a possible acquisition are incorrect" and that the company "has not held discussions with Nvidia about an acquisition," while describing NVIDIA as a business partner and investor.[22] The NVIDIA talks did not result in a deal.[25]

### Nebius talks, layoffs and leadership changes (2026)

On April 9, 2026, Calcalist and Globes, citing The Information, reported that AI21 was in acquisition talks with the AI cloud company Nebius, which earlier in the year had agreed to buy the Israeli startup [Tavily](https://aiwiki.ai/wiki/tavily) for up to $400 million; financial terms of the AI21 talks were not disclosed.[25][26] Calcalist reported that AI21 then employed about 200 people, down from about 260 in 2023, and that its annual revenue was estimated at about $50 million.[25]

On May 18, 2026, AI21 told employees that it would lay off about 110 of its 180 staff, leaving about 70, and refocus on AI agent optimization technology built around Maestro.[27][28] According to Calcalist, the acquisition talks with Nebius were terminated and the two companies signed a commercial partnership agreement instead.[27] AI21 said it would discontinue the sale of standalone models, which it described as not a sufficient source of income on their own, and that it had signed contracts worth tens of millions of dollars with international customers, including Nebius, to deploy Maestro, as well as partnership agreements with companies including Wix.[27][28] In a statement, the company said that developments in AI had "forced us to reexamine the company's operations from end to end."[27] Calcalist reported that AI21 would also halt development of the Jamba model family.[29]

Calcalist further reported, citing unnamed sources, that a significant part of AI21's engineering team was expected to move to Nebius in what appeared to be an acqui-hire-style arrangement worth tens of millions of dollars. AI21 denied that it had been sold and described the deal as a commercial partnership under which Nebius would integrate Maestro into its cloud services. Nebius said: "Nebius has agreed to license AI21's technology and we will work together to productise it in a Nebius context. There is no 'acqui-hire' and claiming otherwise is false and misleading."[29]

On July 7, 2026, Calcalist reported that Shashua had informed the company of his intention to step down from AI21 and its board of directors.[30] In August 2026 Globes wrote that he had recently announced his retirement as chairman of AI21 Labs, at a time when he was also stepping down as chief executive of Mobileye to concentrate on [doubleAI](https://aiwiki.ai/wiki/doubleai), the AI startup (also known as AAI) that he had co-founded in 2023.[31][30] As of October 2026, AI21's website listed Goshen as CEO and Shoham as chairman.[32]

### Shift to post-training (2026)

AI21 changed how it presented its business over the following months. In late May 2026 its website still led with the Jamba models and Maestro.[51] By September 2026 it offered three products aimed at the cost and quality of AI agents: an "Intelligent Gateway" that reduces token spending on agent traffic, a "Harness Optimizer" that searches for the best agent [harness](https://aiwiki.ai/wiki/harness) configuration against a customer's own evaluations, and a post-training service that tunes small open models to match frontier quality on a customer's workloads.[36] By October 2026 the site described AI21 as offering "lab-as-a-service" model training, in which "Forward-Deployed Researchers" post-train specialized models on a customer's data and the customer owns the resulting model.[33]

On October 15, 2026, AI21 said its researchers had spent three weeks post-training NVIDIA's [Nemotron 3.5 Super VL](https://aiwiki.ai/wiki/nemotron_3_5_super_vl) with three customers: the revenue AI company Gong, the HR and payroll platform Deel, and the AI security company Alice. According to AI21, Gong's own evaluations found that its model matched state-of-the-art models such as Claude Opus 5.5 on speed and accuracy at a small fraction of the cost; the Deel model reached 72% accuracy on an eight-country payroll benchmark, 8 points above Claude Opus 5.5 at 33 times lower cost per query; and the Alice model improved its resilience to prompt-injection attacks by 41 points.[34] AI21 also announced plans to attend [NVIDIA GTC Berlin 2026](https://aiwiki.ai/wiki/nvidia_gtc_berlin_2026), held October 20-22.[50]

## Language Models

### Jurassic-1 (2021)

Jurassic-1 (J1) was AI21 Labs' first family of large language models, announced alongside the AI21 Studio developer platform on August 4, 2021.[1] The family consisted of two [autoregressive model](https://aiwiki.ai/wiki/autoregressive_model)s:

- **J1-Jumbo**: 178 billion parameters, making it one of the largest publicly accessible language models at the time of release.
- **J1-Large**: 7 billion parameters, offering a smaller and faster alternative.

A distinguishing feature of Jurassic-1 was its vocabulary of approximately 250,000 tokens, roughly five times larger than most existing vocabularies at the time.[14] This vocabulary included multi-word tokens such as common expressions, phrases, and named entities, which improved computational efficiency, reduced latency, and allowed more text to fit within a fixed context window. The larger vocabulary also meant that Jurassic-1 could pack more examples into a prompt for [few-shot learning](https://aiwiki.ai/wiki/few-shot_learning) compared to similarly sized models like [GPT-3](https://aiwiki.ai/wiki/gpt-3).[14]

Jurassic-1 was trained on a broad corpus spanning web text, academic publications, legal documents, and source code. The architecture diverged from the standard Transformer design used by GPT-3; AI21 Labs optimized the depth-to-width ratio based on theoretical work on expressivity tradeoffs in self-attention networks.[14]

### Jurassic-2 (2023)

Jurassic-2 (J2) was released on March 9, 2023, as the successor to Jurassic-1. The model family offered several improvements over its predecessor, including better quality, lower latency (up to 30% faster inference), multilingual support, and zero-shot instruction-following capabilities.[2]

The Jurassic-2 lineup included three tiers:

- **J2-Ultra** (originally J2-Jumbo): The largest and most capable model in the family.
- **J2-Mid** (originally J2-Grande): A mid-sized model balancing quality and cost.
- **J2-Light** (originally J2-Large): The smallest and fastest model, optimized for cost efficiency.

Both J2-Ultra and J2-Mid were available in instruction-tuned variants that could follow natural language instructions without requiring task-specific [fine-tuning](https://aiwiki.ai/wiki/fine_tuning). Jurassic-2 also added support for multiple European languages, including Spanish, French, German, Portuguese, Italian, and Dutch.[2]

Jurassic-2 models were made available through AI21 Studio, [Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock), and [Amazon SageMaker](https://aiwiki.ai/wiki/amazon_sagemaker), expanding AI21 Labs' reach to enterprise customers already using AWS infrastructure.[2]

### What is Jamba? (March 2024)

Jamba, released on March 28, 2024, represented a fundamental architectural shift for AI21 Labs. Rather than building on the standard Transformer architecture, Jamba introduced a hybrid design combining Transformer attention layers with Mamba structured state space model (SSM) layers, augmented by a [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts) (MoE) mechanism.[3] AI21 Labs described Jamba as "the world's first production-grade Mamba based model," enhancing the SSM with Transformer elements to compensate for the limitations of a pure SSM design.[3]

The Mamba architecture, introduced by [Albert Gu](https://aiwiki.ai/wiki/albert_gu) and [Tri Dao](https://aiwiki.ai/wiki/tri_dao) in December 2023, uses selective state spaces to process sequences in linear time rather than the quadratic time complexity of standard Transformer attention.[7] This makes it particularly efficient for long sequences; according to AI21, no one had scaled Mamba beyond 3 billion parameters before Jamba.[3]

Jamba's architecture interleaves blocks of Transformer and Mamba layers, with each block containing either an attention or a Mamba layer followed by a multi-layer [perceptron](https://aiwiki.ai/wiki/perceptron) (MLP). The overall design uses a ratio of one Transformer attention layer for every eight total layers. MoE is applied to selected layers to increase total [model capacity](https://aiwiki.ai/wiki/model_capacity) while keeping the number of active parameters manageable during inference.[4]

Key specifications of the original Jamba model:

| Specification | Value |
|---|---|
| Total Parameters | 52 billion |
| Active Parameters | 12 billion |
| Context Window | 256K tokens |
| Architecture | Hybrid Mamba-Transformer + MoE |
| License | Apache 2.0 |
| Max Context on Single 80GB GPU | ~140K tokens |

Jamba demonstrated approximately 3x throughput on long contexts compared to [Mixtral](https://aiwiki.ai/wiki/mixtral) 8x7B, a leading open MoE model at the time.[3] It also matched or outperformed comparable models on standard language benchmarks and long-context evaluations. The model was released as open weights on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the Apache 2.0 license, making it freely available for commercial use.[3]

Jamba's release was significant because it demonstrated that the hybrid SSM-Transformer approach could be scaled to production quality, challenging the assumption that pure Transformer architectures were the only viable path for building competitive large language models.

### Jamba 1.5 (August 2024)

On August 22, 2024, AI21 Labs released the Jamba 1.5 model family, scaling the hybrid Mamba-Transformer architecture to larger sizes.[5] The family included two models:

- **Jamba 1.5 Large**: 398 billion total parameters with 94 billion active parameters, representing the first time a hybrid SSM-Transformer model had been scaled to this size.
- **Jamba 1.5 Mini**: 52 billion total parameters with 12 billion active parameters, matching the original Jamba's size but with improved training and instruction-following.

Both models retained the 256K context window and hybrid architecture of the original Jamba. They added support for [function calling](https://aiwiki.ai/wiki/function_calling), [retrieval-augmented generation](https://aiwiki.ai/wiki/retrieval_augmented_generation) (RAG) optimizations, structured JSON output, and multilingual capabilities.[5]

Jamba 1.5 Large was positioned as AI21's most advanced model, capable of handling complex reasoning tasks such as financial analysis and legal document review. Jamba 1.5 Mini was designed for speed and efficiency in tasks like customer support, document summarization, and general text generation.

The Jamba 1.5 models were released as open weights under the Jamba Open Model License and made available through AI21 Studio, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI.[5]

### Jamba 1.6 (March 2025)

Jamba 1.6, released on March 6, 2025, came in Large and Mini sizes with a 256K context window. AI21 marketed it as an open model for private enterprise deployment, downloadable from Hugging Face for on-premises or in-VPC use, and reported that Jamba Large 1.6 outperformed Mistral Large 2, Llama 3.3 70B and Command R+ on quality while Jamba Mini 1.6 outperformed Ministral 8B, Llama 3.1 8B and Command R7B. The release also introduced a Batch API for high-volume asynchronous requests.[37] In June 2025 the company announced that Jamba 1.6 was available as an [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) microservice and on HPE Private Cloud AI.[48]

### Jamba 1.7 (2025)

AI21's documentation dates the API snapshots of Jamba 1.7 Large and Jamba 1.7 Mini to July 2025.[16] The 1.7 generation kept the same parameter counts and 256K context window as Jamba 1.5 (398B total / 94B active for Large; 52B total / 12B active for Mini) and focused on improvements in grounding and instruction-following. Like Jamba 1.5, the 1.7 models officially support nine languages: English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic and Hebrew.[16][5]

### Jamba Reasoning 3B (October 2025)

Jamba Reasoning 3B, released on October 8, 2025, is a 3-billion-parameter open reasoning model under the Apache 2.0 license with a 256K-token context window, which AI21 said could handle up to 1 million tokens. The company said its hybrid design used a KV cache eight times smaller than a standard Transformer's and generated 40 tokens per second on an M3 MacBook Pro at a 32K-token context, and it distributed the model through Hugging Face, Kaggle, LM Studio and llama.cpp for on-device use.[38]

### Jamba 2 (January 2026)

Jamba 2 was released on January 8, 2026, as a pair of compact models focused on enterprise reliability and steerability.[6] The family includes:

- **Jamba 2 3B**: A dense (non-MoE) model with 3 billion parameters.
- **Jamba 2 Mini**: A MoE model with 52 billion total parameters and 12 billion active parameters.

Both models feature a 256K context window and were released under the Apache 2.0 license. Jamba 2 was mid-trained on 500 billion carefully curated tokens with increased representation of math, code, high-quality web data, and long documents. The training process included a state passing phase for the Mamba layers, cold-start [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning), [direct preference optimization (DPO)](https://aiwiki.ai/wiki/direct_preference_optimization_dpo), and multi-phase on-policy [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning).[6]

The Jamba 2 models were designed for enterprise workflows that demand accuracy and grounding, producing answers that stay closely tied to source material across technical manuals, research papers, company policies, and internal knowledge bases. In benchmarks, Jamba 2 led on instruction-following tests ([IFBench](https://aiwiki.ai/wiki/ifbench), IFEval, Collie) and grounding evaluations (FACTS).[6]

In May 2026 AI21 said it would stop selling standalone models, and Calcalist reported that the company would halt development of the Jamba family.[28][29] As of October 2026, Jamba 2 was the most recent Jamba generation AI21 had released.[49]

### Model Comparison Table

| Model | Release Date | Total Parameters | Active Parameters | Context Window | Architecture | License |
|---|---|---|---|---|---|---|
| Jurassic-1 Jumbo | August 2021 | 178B | 178B | 2,048 tokens | Transformer | Proprietary |
| Jurassic-1 Large | August 2021 | 7B | 7B | 2,048 tokens | Transformer | Proprietary |
| Jurassic-2 Ultra | March 2023 | Undisclosed | Undisclosed | 8,192 tokens | Transformer | Proprietary |
| Jurassic-2 Mid | March 2023 | Undisclosed | Undisclosed | 8,192 tokens | Transformer | Proprietary |
| Jurassic-2 Light | March 2023 | Undisclosed | Undisclosed | 8,192 tokens | Transformer | Proprietary |
| Jamba | March 2024 | 52B | 12B | 256K tokens | Hybrid Mamba-Transformer + MoE | Apache 2.0 |
| Jamba 1.5 Mini | August 2024 | 52B | 12B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba 1.5 Large | August 2024 | 398B | 94B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba 1.6 Mini | March 2025 | 52B | 12B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba 1.6 Large | March 2025 | 398B | 94B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba 1.7 Mini | July 2025 | 52B | 12B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba 1.7 Large | July 2025 | 398B | 94B | 256K tokens | Hybrid Mamba-Transformer + MoE | Jamba Open Model License |
| Jamba Reasoning 3B | October 2025 | 3B | 3B | 256K tokens | Hybrid Mamba-Transformer | Apache 2.0 |
| Jamba 2 3B | January 2026 | 3B | 3B | 256K tokens | Hybrid Mamba-Transformer (dense) | Apache 2.0 |
| Jamba 2 Mini | January 2026 | 52B | 12B | 256K tokens | Hybrid Mamba-Transformer + MoE | Apache 2.0 |

## Products and Services

### Maestro: AI planning and orchestration

Maestro is AI21 Labs' AI planning and orchestration system for the enterprise, introduced on March 10, 2025 and described by the company as "the world's first AI Planning and Orchestration System built for the enterprise."[19] Rather than being a single model, Maestro is a framework (which AI21 calls the AI Planning and Orchestration System, or AIPOS) that wraps around large language models to analyze a task, plan a multi-step solution, execute it, and validate the result against requirements.[20]

AI21 Labs reported that Maestro boosts the instruction-following accuracy of paired models by up to 50%, raising the accuracy of models such as [GPT-4o](https://aiwiki.ai/wiki/gpt_4o) and [Claude 3.5 Sonnet](https://aiwiki.ai/wiki/claude_3_5_sonnet), and enabling [reasoning models](https://aiwiki.ai/wiki/reasoning_models) like o3-mini to exceed 95% accuracy on a benchmark of generation tasks with multiple requirements.[19][20] The company framed the system as a response to the gap between AI demos and production deployment, citing an AWS finding that only 6% of organizations had a [generative AI](https://aiwiki.ai/wiki/generative_ai) application in deployment.[20] Maestro was made available initially through early access, with broader SaaS and virtual private cloud (VPC) availability planned for later in 2025.[20]

AI21 made Maestro available to enterprise customers in July 2025,[39] and in December 2025 announced general availability for deployment in Amazon virtual private clouds alongside models served through Amazon Bedrock.[40] Globes described Maestro as the product that had grown fastest, and Calcalist described it as the company's only remaining active product line after the May 2026 restructuring.[21][29] By October 2026, AI21's website no longer listed Maestro among its products.[33]

### Post-training service

AI21's post-training offering, which the company calls "lab-as-a-service", pairs customers with its Forward-Deployed Researchers, who build evaluation suites from the customer's tasks, post-train a model and its agent harness together, and optimize the result for serving either by AI21 or inside the customer's own environment. AI21 says the customer keeps the evaluation suite and training recipe and owns the resulting model, and that training, evaluations and serving can run in the customer's VPC or on premises.[33] Its first published results, in October 2026, used NVIDIA's open Nemotron 3.5 Super VL as the base model.[34]

### AI21 Gateway

AI21 Gateway is a drop-in API endpoint that sits between an agent and its models and tries to cut token spending through per-agent cost profiles, routing of individual turns to different models, context compaction and caching; AI21 advertises savings of up to 40%. The company says it works with off-the-shelf agents such as Claude Code and Codex, with frameworks such as LangChain, and with custom agents, and that it checks savings against coding benchmarks to preserve quality. As of October 2026 the gateway was in beta and free to use during the preview.[35]

### AI21 Studio

AI21 Studio is the company's developer platform, providing API access to AI21's language models and task-specific capabilities. Launched alongside Jurassic-1 in August 2021, the platform allows developers to build text-based applications including virtual assistants, [chatbot](https://aiwiki.ai/wiki/chatbot)s, content generation tools, text classifiers, and more.[1]

The platform offers several access tiers and includes features such as custom model training (requiring as few as 50 to 100 training examples for fine-tuning, according to the launch announcement[1]), a web-based interactive playground for testing prompts, and comprehensive API documentation. AI21 Studio supports both the Jurassic and Jamba model families.

### Task-Specific APIs

Alongside Jurassic-2, AI21 Labs introduced a set of Task-Specific Models (TSMs): pre-built, highly optimized APIs designed for specific natural language processing tasks. Unlike general-purpose language model APIs, these endpoints require no [prompt engineering](https://aiwiki.ai/wiki/prompt_engineering) or fine-tuning and can be integrated with a single API call.[2]

The Task-Specific APIs include:

| API | Function |
|---|---|
| Paraphrase | Generates alternative phrasings of input text while preserving meaning, with adjustable tone and length |
| Summarize | Condenses long documents into concise summaries, optimized for financial reports, legal documents, and technical papers |
| Contextual Answers | Answers user questions based on a provided knowledge base or document context |
| Grammatical Error Correction | Identifies and corrects grammatical errors in text |
| Text Improvements | Suggests improvements to text for clarity, fluency, and engagement |

AI21 Labs has reported that its Summarize API achieved a faithfulness rate 19% higher than OpenAI's Davinci-003 model and an acceptance rate 18% higher on the same benchmark.[2]

### Wordtune

Wordtune is AI21 Labs' consumer-facing AI writing assistant, launched on October 27, 2020.[11] The product was the company's first public offering and served as the vehicle through which AI21 emerged from stealth.[11]

Wordtune is available as Chrome and Microsoft Edge browser extensions, a web application, and an iOS app.[47] Its main features include:

- **Rewriting and paraphrasing**: Users can highlight text and receive multiple alternative phrasings that preserve the original meaning.[43]
- **Tone adjustment**: The tool can make text more formal or more casual, and make sentences shorter or longer.[43]
- **Wordtune Spices**: Launched on January 17, 2023, Spices lets writers request generated additions to their own text, such as an expansion on a topic, a counterargument, an analogy, a statistical or historical fact, a quote, or a joke, and in some cases links to a web page as the source of the information.[46]
- **Wordtune Read**: Launched in November 2021, this feature summarizes documents and web articles. Users upload a PDF or paste a link, and a Spotlight feature re-summarizes the text from different points of view.[15]

Google chose Wordtune as one of its favorite Chrome extensions of 2021, and in July 2022 AI21 said the extension had millions of active users.[10] As of October 2026 the Wordtune website claimed more than 10 million users and offered a free tier alongside paid subscriptions.[47] According to Calcalist, AI21 halted development of Wordtune in April 2025 to concentrate on enterprise customers.[12][29]

### Cloud Platform Integrations

AI21 Labs models have been available through multiple cloud platforms, broadening access for enterprise customers:[2][5][41][48]

| Platform | Available models or products |
|---|---|
| [Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock) | Jurassic-2, Jamba 1.5 |
| [Amazon SageMaker](https://aiwiki.ai/wiki/amazon_sagemaker) | Jurassic-2 |
| [Google Cloud Vertex AI](https://aiwiki.ai/wiki/google_vertex_ai) | Jamba 1.5 |
| Microsoft Azure AI | Jamba 1.5 |
| NVIDIA API Catalog | Jamba |
| [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) | Jamba 1.6 |
| HPE Private Cloud AI | Jamba 1.6 |
| [Together AI](https://aiwiki.ai/wiki/together_ai) | Maestro, connected to Together-hosted open models |

These integrations allow enterprises to deploy AI21 models within their existing cloud infrastructure without managing separate model hosting. In May 2026, AI21 said it would discontinue the sale of standalone models.[28]

## How does the hybrid SSM-Transformer architecture work?

The most distinctive technical contribution from AI21 Labs is the hybrid Mamba-Transformer architecture used in the Jamba model family. Understanding this approach requires some background on the two architectures it combines.

### Transformer Attention

The standard Transformer architecture, introduced in 2017, uses self-attention mechanisms to process input sequences. [Self-attention](https://aiwiki.ai/wiki/self_attention) allows every token in a sequence to attend to every other token, capturing long-range dependencies effectively. However, the computational cost of self-attention scales quadratically with sequence length (O(n^2)), which makes processing very long sequences expensive in terms of both time and memory.

### State Space Models and Mamba

Structured state space models (SSMs) offer an alternative approach to sequence modeling. SSMs process sequences by maintaining a hidden state that is updated as each new token arrives, similar to [recurrent neural networks](https://aiwiki.ai/wiki/rnn) but with a mathematically structured state transition. This allows SSMs to process sequences in linear time (O(n)), making them far more efficient for long inputs.

The Mamba architecture, introduced by Albert Gu and Tri Dao in their December 2023 paper "Mamba: Linear-Time Sequence Modeling with Selective State Spaces," improved upon earlier SSMs by adding a selective mechanism that allows the model to focus on relevant parts of the input.[7] This selectivity addressed a key weakness of previous SSMs, which struggled with content-based reasoning tasks.[7]

### How Jamba Combines Both

Jamba's architecture interleaves Mamba and Transformer layers in a structured pattern. Each block contains either a Mamba SSM layer or a Transformer attention layer, followed by a standard multi-layer perceptron. The ratio of Mamba layers to Transformer layers is approximately 7:1, meaning that only one out of every eight layers uses the computationally expensive attention mechanism.[4]

This hybrid approach offers several benefits:

1. **Memory efficiency**: The Mamba layers maintain a compact recurrent state rather than storing full key-value caches for all past tokens, dramatically reducing memory requirements for long contexts.
2. **Throughput**: The linear-time processing of Mamba layers delivers higher throughput on long sequences compared to pure Transformer models.
3. **Quality preservation**: The interspersed Transformer attention layers provide the content-based reasoning and long-range dependency modeling capabilities that pure SSM models can lack.
4. **MoE scaling**: The Mixture of Experts mechanism allows the model to have a large total parameter count while keeping the active parameter count (and thus inference cost) low.

The result is a model that can process 256K token contexts efficiently on standard hardware, fitting up to 140K tokens on a single 80GB GPU, something that would require significantly more resources with a pure Transformer model of comparable quality.[4]

## How much funding has AI21 Labs raised?

AI21 Labs has raised $336 million in total funding through its seed, Series A, Series B, and Series C rounds and other investments, with its last completed raise in November 2023.[9] A $300 million round backed by Google and NVIDIA, widely reported in 2025, was never closed.[17][22]

| Round | Date | Amount | Valuation | Investors named |
|---|---|---|---|---|
| Seed | 2019 | $9.5M | Undisclosed | Same investors as the Series A, per Calcalist |
| Series A | November 2020 | $25M | Undisclosed | Pitango First (lead), TPY Capital |
| Walden Catalyst investment | November 2021 | $20M | $400M (reported) | Walden Catalyst |
| Series B | July 2022 | $64M | $664M | Ahren Innovation Capital (lead), Amnon Shashua, Walden Catalyst, Pitango, TPY Capital, Mark Leslie |
| Series C (Tranche 1) | August 2023 | $155M | $1.4B | Walden Catalyst, Pitango, SCB10X, b2venture, Samsung Next, Amnon Shashua; Google and NVIDIA participated |
| Series C (Tranche 2) | November 2023 | $53M | $1.4B | Intel Capital, Comcast Ventures, Ahren Innovation Capital |
| Reported Series D | 2025 | $300M (reported) | Not disclosed | Not completed, according to Calcalist and Globes |
| **Total raised** | | **$336M** | **$1.4B (last priced round)** | |

The seed, Series A and Walden Catalyst figures come from Calcalist;[43][44] the Series B and Series C figures come from the company's announcements and TechCrunch,[10][8][9] and the status of the reported Series D from Calcalist and Globes.[17][22] Cumulative totals are as reported by the company: $118 million after the Series B, $283 million after the first Series C tranche, and $336 million after the second.[10][9]

Notable investors across all rounds include Google, NVIDIA, Intel Capital, Comcast Ventures, Pitango, Walden Catalyst, Ahren Innovation Capital, TPY Capital, Samsung Next, SCB10X, b2venture, and co-founder Amnon Shashua.[8][9][10] Calcalist has also listed 8VC and Coatue among the company's investors.[29]

## How does AI21 Labs compare to its competitors?

Before its 2026 restructuring, AI21 Labs competed in a crowded market for large language models and enterprise AI services. Its primary competitors included several well-funded companies:

| Company | Headquarters | Key Models | Focus |
|---|---|---|---|
| [OpenAI](https://aiwiki.ai/wiki/openai) | San Francisco, USA | [GPT-4](https://aiwiki.ai/wiki/gpt-4), [GPT-4o](https://aiwiki.ai/wiki/gpt_4o) | General-purpose AI, consumer and enterprise |
| [Anthropic](https://aiwiki.ai/wiki/anthropic) | San Francisco, USA | [Claude](https://aiwiki.ai/wiki/claude) | [AI safety](https://aiwiki.ai/wiki/ai_safety), enterprise |
| [Cohere](https://aiwiki.ai/wiki/cohere) | Toronto, Canada | [Command R](https://aiwiki.ai/wiki/command_r) | Enterprise NLP |
| [Mistral AI](https://aiwiki.ai/wiki/mistral_ai) | Paris, France | Mistral, Mixtral | Open-weight models, enterprise |
| [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) | London, UK | [Gemini](https://aiwiki.ai/wiki/gemini) | General-purpose AI, research |
| [Meta AI](https://aiwiki.ai/wiki/meta_ai) | Menlo Park, USA | [LLaMA](https://aiwiki.ai/wiki/llama) | Open-source models |

AI21 Labs sought to differentiate itself in several ways. First, its hybrid Mamba-Transformer architecture gave the Jamba models an efficiency advantage for long-context processing compared to pure Transformer competitors. Second, the company maintained a strong enterprise focus, offering deployment flexibility through SaaS, cloud partnerships (AWS, Google Cloud, Azure), virtual private cloud (VPC), and on-premises options. Third, its Task-Specific APIs and the Maestro orchestration layer provided specialized, ready-to-use solutions that emphasized accuracy and reliability.

In a September 2023 interview with VentureBeat, co-founder Yoav Shoham said that OpenAI was the main competitor AI21 met in enterprise deals and that "when we're invited to the party, we usually win"; he added that when AI21 lost deals it was mostly to OpenAI, sometimes because of brand recognition.[13]

Compared to open-source competitors like Meta's LLaMA and [Mistral AI](https://aiwiki.ai/wiki/mistral_ai)'s models, AI21 Labs occupied a middle ground: the Jamba base models were released with open weights, while the company also offered access through its commercial API platform.

Commentators have been critical of the company's trajectory. In May 2026 Calcalist wrote that AI21 had failed to complete another major funding round after 2023 while competitors raised large sums, and that industry estimates put its annual revenue below $50 million for several years, compared with reported annualized run rates of about $400 million at Mistral and $200 million at Cohere. AI21 attributed some of its fundraising difficulty to the war in Israel and said it had sufficient cash reserves.[29]

## Company Profile

| Detail | Information |
|---|---|
| Founded | 2017 |
| Headquarters | Tel Aviv, Israel |
| Co-Founders | [Yoav Shoham](https://aiwiki.ai/wiki/yoav_shoham), Ori Goshen, [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua) |
| Leadership (October 2026) | Ori Goshen (CEO), Yoav Shoham (chairman) |
| Employees | About 70 after the May 2026 restructuring (down from about 180) |
| Total Funding | $336 million |
| Valuation | $1.4 billion (Series C, 2023; no later priced round completed) |
| Products (October 2026) | Post-training service, AI21 Gateway (beta) |
| Earlier products | Jurassic and Jamba models, AI21 Studio, Maestro, [Wordtune](https://aiwiki.ai/wiki/wordtune), Task-Specific APIs |

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Jamba](https://aiwiki.ai/wiki/jamba)
- [Jamba 2](https://aiwiki.ai/wiki/jamba_2)
- [Jamba Reasoning 3B](https://aiwiki.ai/wiki/jamba_reasoning_3b)
- [Wordtune](https://aiwiki.ai/wiki/wordtune)
- [Mamba](https://aiwiki.ai/wiki/mamba)
- [Nebius](https://aiwiki.ai/wiki/nebius)
- [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua)
- [Yoav Shoham](https://aiwiki.ai/wiki/yoav_shoham)

## References

1. AI21 Labs. "[Announcing AI21 Studio and Jurassic-1 Language Models](https://www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1/)." AI21 Blog, August 4, 2021.
2. AI21 Labs. "[Announcing Jurassic-2 and Task-Specific APIs](https://www.ai21.com/blog/introducing-j2/)." AI21 Blog, March 9, 2023.
3. AI21 Labs. "[Introducing Jamba: AI21's Groundbreaking SSM-Transformer Model](https://www.ai21.com/blog/announcing-jamba/)." AI21 Blog, March 28, 2024.
4. Lieber, Opher, et al. "[Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887)." arXiv:2403.19887, March 2024.
5. AI21 Labs. "[The Jamba 1.5 Open Model Family: The Most Powerful and Efficient Long Context Models](https://www.ai21.com/blog/announcing-jamba-model-family/)." AI21 Blog, August 22, 2024.
6. AI21 Labs. "[Introducing Jamba2: The Open Source Model Family for Enterprise Reliability and Efficiency](https://www.ai21.com/blog/introducing-jamba2/)." AI21 Blog, January 8, 2026.
7. Gu, Albert and Tri Dao. "[Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://arxiv.org/abs/2312.00752)." arXiv:2312.00752, December 2023.
8. TechCrunch. "[Generative AI startup AI21 Labs lands $155M at a $1.4B valuation](https://techcrunch.com/2023/08/30/generative-ai-startup-ai21-labs-lands-155m-at-a-1-4b-valuation/)." August 30, 2023.
9. AI21 Labs. "[AI21 Completes $208 Million Oversubscribed Series C Round](https://www.ai21.com/blog/ai21-completes-208-million-oversubscribed-series-c-round/)." Press Release, November 21, 2023.
10. PR Newswire. "[AI21 Labs Raises $64 Million to Change the Way People Read and Write Using Artificial Intelligence](https://www.prnewswire.com/news-releases/ai21-labs-raises-64-million-to-change-the-way-people-read-and-write-using-artificial-intelligence-301584831.html)." July 12, 2022.
11. BusinessWire. "[AI21 Labs Comes out of Stealth and Launches First Deep-Tech Writing Assistant, Wordtune](https://www.businesswire.com/news/home/20201027005162/en/AI21-Labs-Comes-out-of-Stealth-and-Launches-First-Deep-Tech-Writing-Assistant-Wordtune)." October 27, 2020.
12. Sophie Shulman and Meir Orbach, Calcalist Tech. "[Nvidia in advanced talks to acquire AI21 in $2-3 billion deal focused on talent](https://www.calcalistech.com/ctechnews/article/rkbh00xnzl)." December 30, 2025.
13. Sharon Goldman, VentureBeat. "[AI21 Labs co-founder says 'we usually win' when competing with OpenAI for enterprise business](https://venturebeat.com/ai/ai21-labs-co-founder-says-we-usually-win-when-competing-with-openai-for-enterprise-business)." September 5, 2023.
14. Lieber, Opher, et al. "[Jurassic-1: Technical Details and Evaluation](https://www.ai21.com/research/jurassic-1-technical-details-evaluation/)." AI21 Labs White Paper, 2021.
15. AI21 Labs. "[Introducing Wordtune Read](https://www.ai21.com/blog/introducing-wordtune-read/)." AI21 Blog, November 14, 2021.
16. AI21 Labs. "[Jamba Foundation Models](https://docs.ai21.com/docs/jamba-foundation-models)." AI21 Documentation, accessed October 2026.
17. Allon Sinai, Calcalist Tech. "[AI21 never closed its reported $300 million round as Nvidia weighs an acquisition](https://www.calcalistech.com/ctechnews/article/rjswz37eze)." January 1, 2026.
18. Calcalist Tech. "['There are talks with Nvidia and others': Shashua confirms AI21 acquisition discussions](https://www.calcalistech.com/ctechnews/article/h1e8rttn11x)." January 8, 2026.
19. PR Newswire. "[AI21 Introduces Maestro, the World's First AI Planning and Orchestration System Built for the Enterprise](https://www.prnewswire.com/news-releases/ai21-introduces-maestro-the-worlds-first-ai-planning-and-orchestration-system-built-for-the-enterprise-302397075.html)." March 10, 2025.
20. AI21 Labs. "[Meet Maestro: The AI System for Automating Data-Intensive Enterprise Tasks](https://www.ai21.com/blog/maestro-ai-planning-orchestration/)." AI21 Blog, March 10, 2025.
21. Assaf Gilead, Globes. "[After raising $700m, AI21 Labs looks for exit](https://en.globes.co.il/en/article-after-raising-700m-ai21-labs-looks-for-exit-1001530580)." December 30, 2025.
22. Assaf Gilead, Globes. "[AI21 Labs denies Nvidia acquisition talks](https://en.globes.co.il/en/article-ai21-labs-denies-nvidia-acquisition-talks-1001531815)." January 12, 2026.
23. Shubhendu Vimal, Verdict (via Yahoo Finance). "[Nvidia in talks to buy Israel's AI21 Labs for up to $3bn - report](https://finance.yahoo.com/news/nvidia-talks-buy-israel-ai21-111146958.html)." December 31, 2025.
24. Maria Deutscher, SiliconANGLE. "[Report: Nvidia could acquire LLM startup AI21 Labs for $3B](https://siliconangle.com/2025/12/30/report-nvidia-acquire-llm-startup-ai21-labs-3b/)." December 30, 2025.
25. Calcalist Tech. "[Nebius in talks to acquire AI21 after Nvidia deal falls through](https://www.calcalistech.com/ctechnews/article/r1kdh64nzg)." April 9, 2026.
26. Assaf Gilead, Globes. "[Nebius in talks to acquire AI21 Labs - report](https://en.globes.co.il/en/article-1001539693)." April 9, 2026.
27. Calcalist Tech. "[AI21 cuts more than 60% of its workforce in major strategic overhaul](https://www.calcalistech.com/ctechnews/article/rjwumhukfx)." May 18, 2026.
28. Assaf Gilead, Globes. "[Shashua's AI21 Labs laying off 60% of employees](https://en.globes.co.il/en/article-shashuas-ai21-labs-laying-off-60-of-employees-1001543241)." May 18, 2026.
29. Sophie Shulman, Calcalist Tech. "[AI21's fall from unicorn to survival mode](https://www.calcalistech.com/ctechnews/article/by34n00tjfl)." May 19, 2026.
30. Meir Orbach, Calcalist Tech. "[Amnon Shashua to leave AI21 after nine years](https://www.calcalistech.com/ctechnews/article/r1eabd5qfl)." July 7, 2026.
31. Globes. "[Amnon Shashua doubles down on doubleAI](https://en.globes.co.il/en/article-amnon-shashua-doubles-down-on-doubleai-1001551128)." August 2, 2026.
32. AI21 Labs. "[About](https://www.ai21.com/about/)." Accessed October 27, 2026.
33. AI21 Labs. "[Post-training for specialized intelligence (homepage)](https://www.ai21.com/)." Accessed October 27, 2026.
34. AI21 Labs. "[AI21 Partners with Gong, Deel and Alice to Post-Train Specialized Models on NVIDIA Nemotron 3.5 Super VL](https://www.ai21.com/blog/ai21-post-trains-nvidia-nemotron-35-super-vl/)." AI21 Blog, October 15, 2026.
35. AI21 Labs. "[AI21 Gateway](https://www.ai21.com/gateway/)." Accessed October 27, 2026.
36. Internet Archive. "[Snapshot of ai21.com, September 11, 2026](https://web.archive.org/web/20260911092728/https://www.ai21.com/)." Archived September 11, 2026.
37. AI21 Labs. "[AI21's Jamba 1.6: The Best Open Model for Private Enterprise Deployment](https://www.ai21.com/blog/introducing-jamba-1-6/)." AI21 Blog, March 6, 2025.
38. AI21 Labs. "[Introducing Jamba Reasoning 3B: Tiny Model, Huge Possibilities](https://www.ai21.com/blog/introducing-jamba-reasoning-3b/)." AI21 Blog, October 8, 2025.
39. AI21 Labs. "[AI21 Maestro: AI Built for Enterprise Knowledge Work](https://www.ai21.com/blog/maestro-ai-enterprise-knowledge-work/)." AI21 Blog, July 8, 2025.
40. AI21 Labs, ACCESS Newswire (via Yahoo Finance). "[AI21 Launches Maestro on Amazon VPC, Bringing Enterprise-Grade Agentic AI to Every Organization](https://finance.yahoo.com/news/ai21-launches-maestro-amazon-vpc-204500207.html)." December 1, 2025.
41. AI21 Labs, ACCESS Newswire (via Yahoo Finance). "[AI21 Labs and Together AI Partner to Expand Access to Open-Source Models](https://finance.yahoo.com/news/ai21-labs-together-ai-partner-140000030.html)." November 26, 2025.
42. AI21 Labs, ACCESS Newswire (via Yahoo Finance). "[AI21 Labs and Fnac Darty, Partner to Drive Retail AI Transformation](https://uk.finance.yahoo.com/news/ai21-labs-fnac-darty-partner-080000525.html)." December 8, 2025.
43. Meir Orbach, Calcalist Tech. "[AI21 Labs completes $25 million round led by Pitango](https://www.calcalistech.com/articles/0,7340,L-3874192,00.html)." November 19, 2020.
44. Meir Orbach, Calcalist Tech. "[Amnon Shashua's A21 Labs raises $20 million from Walden Catalyst at $400 million valuation](https://www.calcalistech.com/ctech/articles/0,7340,L-3921431,00.html)." November 1, 2021.
45. Globes. "[AI21 Labs raising $300m - report](https://en.globes.co.il/en/article-a121-labs-raising-300m-report-1001509980)." May 11, 2025.
46. Katyanna Quach, The Register. "[Writing tool from AI21 Labs won't do all the hard work for you](https://www.theregister.com/2023/01/17/spices_ai_writing_tool/)." January 17, 2023.
47. Wordtune. "[Wordtune: Express yourself with confidence](https://www.wordtune.com/)." Accessed October 27, 2026.
48. AI21 Labs. "[Newsroom](https://www.ai21.com/newsroom/)." Accessed October 27, 2026.
49. Hugging Face. "[ai21labs model repositories](https://huggingface.co/ai21labs)." Accessed October 27, 2026.
50. AI21 Labs. "[AI21 at NVIDIA GTC Berlin 2026](https://www.ai21.com/events/gtc-berlin-2026/)." Accessed October 27, 2026.
51. Internet Archive. "[Snapshot of ai21.com, May 25, 2026](https://web.archive.org/web/20260525200016/https://www.ai21.com/)." Archived May 25, 2026.

