What does AI21 Labs do? In October 2026, AI21 Labs is a Tel Aviv AI lab that trains specialized models for individual companies: its own researchers work with a customer to post train an open model on that customer's data and tasks, and the customer keeps the weights, the training recipe and the evaluation suite. Before that, AI21 was best known for building its own large language models (Jurassic, then Jamba) and the Wordtune writing app. It is a real lab with real research history, but it is no longer trying to sell a general chatbot or a flagship model to compete with OpenAI, Anthropic or Google.
This guide explains what the company built, what changed in 2025 and 2026, what it sells today, and when it makes sense to talk to it.
AI21 Labs in one table
| Question | Short answer |
|---|---|
| What is it? | An Israeli AI lab, founded in 2017 |
| Founders | Yoav Shoham, Ori Goshen and Amnon Shashua |
| Best known for | Jurassic models (2021), Jamba models (2024), Wordtune |
| Main business now | Training custom models for enterprises on open foundations |
| Consumer product? | No current consumer chatbot |
| Funding raised | About $336 million, per press reports and AI21's own boilerplate |
| Directory page | AI21 Labs on intelsi.si |
AI21 Labs was founded in 2017 by Prof. Yoav Shoham, Ori Goshen and Prof. Amnon Shashua. The company's About page lists Goshen as CEO and cofounder and Shoham as chairman and cofounder. Calcalist reported in July 2026 that Shashua, also the founder of Mobileye, planned to leave the company and its board; he is no longer listed on the About page.
What AI21 Labs does today
AI21 describes itself as a lab offering "lab as a service for specialized intelligence." In plain terms, it does three things for a paying customer:
- Builds an evaluation suite from the customer's real tasks, so both sides agree on what "better" means for quality, cost and speed.
- Post trains an open model (and tunes the agent setup around it, which AI21 calls the harness) using the customer's data, tools and policies.
- Ships the result either as a service run by AI21 or deployed inside the customer's own cloud account or data center.
The AI21 homepage frames the pitch as "rented AI" versus "owned AI": a general model you rent improves only when its vendor ships an update, while a model trained on your own data is an asset you control. AI21 says training, evaluation and serving can run inside the customer's private cloud or on premises, with scoped and logged access for its staff, and that it keeps no copies of customer data.
The people doing the work are what AI21 calls forward deployed researchers, a small team that embeds with the customer for a few weeks. AI21's job listings describe the company as "the only Israeli foundation lab that have trained LLMs from scratch" and say it now delivers models "with the weights, recipes and evaluations" so customers keep them.
A real example from October 2026
On October 15, 2026, AI21 published a case study on post training NVIDIA's open Nemotron 3.5 Super VL model with three companies over about three weeks:
- Gong, a revenue software company, wanted a cheaper model for an agent that pulls deal details out of calls and emails.
- Deel, the payroll and HR platform, wanted complete answers to payroll questions across several countries.
- Alice, an AI security company, wanted a model that resists indirect prompt injection in Slack and GitHub.
For the Alice project, both AI21 and Alice's partner Inflection AI report that the post trained model's resilience to indirect prompt injection rose by 41 points. These are the companies' own reported results, not independent tests, so treat them as claims rather than settled benchmarks.
The detail that matters for most readers: AI21 is now training on top of someone else's open model (NVIDIA's), not only its own Jamba family. That is the clearest sign of how much the company has changed.
How AI21 got here: a short history
| Year | What happened |
|---|---|
| 2017 | AI21 Labs founded in Tel Aviv |
| 2021 | AI21 Studio and the Jurassic-1 models launch (Jumbo had 178 billion parameters) |
| 2024 | Jamba launches on March 28 with open weights under Apache 2.0 |
| 2025 | Wordtune development stops in April; Maestro is announced on March 10 |
| 2026 | Jamba2 Mini and Jamba2 3B ship on January 8; in May the company cuts most of its staff and stops selling standalone models |
| Oct 2026 | Post training work on NVIDIA Nemotron 3.5 Super VL for Gong, Deel and Alice |
Jurassic: AI21's first big models
On August 4, 2021, AI21 launched AI21 Studio and Jurassic-1, a developer platform with an API and a web playground. The larger Jurassic-1 Jumbo model had 178 billion parameters, which the company's press release noted was slightly larger than OpenAI's GPT-3. Unlike GPT-3 at the time, anyone could sign up for the open beta. This is why searches like "AI21 Labs Jurassic" and "Jurassic-2" still show up: for a while AI21 was one of the few companies outside the big US labs offering a large general model by API.
Jamba: the hybrid architecture
On March 28, 2024, AI21 announced Jamba, which it called the first production grade model based on Mamba. Mamba is a state space model, a different way of processing long sequences than the attention layers in a standard Transformer. Jamba mixes Transformer, Mamba and mixture of experts layers, and the Jamba paper describes it as the first production grade hybrid of attention and state space layers. The first Jamba had 52 billion parameters with 12 billion active at a time, a context window of 256,000 tokens, and open weights under the Apache 2.0 license.
AI21's model documentation still lists three Jamba models:
| Model | Size | Context | Hosted endpoint |
|---|---|---|---|
| Jamba Large 1.7 (July 2025) | 398B parameters, 94B active | 256K tokens | jamba-large |
| Jamba2 Mini (January 2026) | 52B parameters, 12B active | 256K tokens | jamba-mini |
| Jamba2 3B (January 2026) | 3B parameters | 256K tokens | None |
The docs list nine supported languages, text in and text out, and a training knowledge cutoff in 2024. If you are interested in downloading and running Jamba yourself, that is an open weight topic covered elsewhere; this page sticks to the company.
Maestro and the 2026 reset
On March 10, 2025, AI21 introduced Maestro, which it called an AI planning and orchestration system. SiliconANGLE's launch coverage describes how it worked: you give it a task plus requirements and a budget, it breaks the job into steps, tries different combinations of models and tools, and checks the output against your requirements.
Then came the reset. On May 18, 2026, Calcalist and Globes reported that AI21 would cut about 110 of its roughly 180 employees, focus on agent optimization built around Maestro, and stop selling its language models as standalone products. AI21's own statement, as quoted by Globes, said the models were "a major infrastructure for its expertise" but "not a sufficient source of income on their own." Acquisition talks with the cloud company Nebius ended in a commercial partnership instead. A few months earlier, in April 2025, AI21 had stopped developing Wordtune, its writing assistant, according to Calcalist and the Israeli tech site Geektime.
By October 2026 the public pitch had moved again, from Maestro by name to custom post training. The homepage no longer leads with Maestro; it leads with forward deployed researchers and a "harness optimizer" that tunes the model and the agent setup together.
Is AI21 Labs a frontier lab?
Not in the sense people usually mean. Frontier labs such as Google DeepMind, OpenAI and Anthropic train the most capable general models and sell them to millions of people through apps like Gemini, ChatGPT and Claude. AI21 did train its own large models from scratch, and its Jamba architecture work was genuinely new, but it has never had a mass market assistant and it has now stepped back from selling general models.
A closer comparison is Cohere, another lab that focuses on businesses rather than consumers. The difference is that AI21's current offer is a service built around training a model you own, while most enterprise model vendors still sell API access to their own models.
If you want to see where AI21's models sit among everyone else's over time, Epoch AI tracks notable AI models and the compute used to train them in its public data hub.
Who should care, and what to check first
You might talk to AI21 if you run a high volume, repetitive task (millions of similar queries a month), you have your own data, and you want a cheaper model you control instead of paying a frontier API for every call. That is the exact profile its October case study describes.
You probably do not need AI21 if you want a chat assistant for your team, a general API to start a prototype, or a free option. For those, see our comparison of ChatGPT, Claude, Gemini and Grok or our list of free OpenAI API alternatives.
Before you build on Jamba's hosted endpoints, confirm they are still sold to new customers. The documentation still lists them, but the company said in May 2026 that it would stop selling standalone models. Ask in writing.
Pros and cons of working with AI21
Pros
- Long track record: it has trained large models since 2021 and designed the Jamba hybrid architecture.
- You keep the weights, training recipe and evaluation suite, according to AI21.
- Can run inside your own cloud or on premises, which suits regulated data.
- Lists SOC 2 compliance and ISO 27001, 27017 and 27018 certifications in its docs.
Cons
- A much smaller company after the May 2026 layoffs, so ask about support and continuity.
- Strategy has shifted three times in about eighteen months (models, then Maestro, then custom post training).
- Customer results so far are self reported, not independently tested.
- No published price list for its training service; you have to contact sales.
FAQ
What does AI21 Labs do in simple terms?
It is an AI lab that, as of October 2026, trains custom versions of open models for businesses, using each customer's data, and hands over the weights and evaluations. It used to sell its own Jurassic and Jamba models by API.
Who owns AI21 Labs?
It is a private company. Its founders are Yoav Shoham, Ori Goshen and Amnon Shashua, and press reports put total funding at about $336 million from investors including Google. Talks to sell the company to Nebius ended in May 2026 with a commercial partnership instead of a sale.
What is the difference between AI21 Jurassic and Jamba?
Jurassic was AI21's first public model family, launched in 2021 through the AI21 Studio API. Jamba, launched in 2024, mixes Transformer layers with Mamba state space layers to handle long inputs (up to 256,000 tokens) more efficiently, and its weights are openly available.
Is AI21 Labs still making Jamba models?
AI21 said in May 2026 that it would stop selling standalone models, and Calcalist reported that Jamba development would halt. The docs still list Jamba Large 1.7 and Jamba2 Mini endpoints, so check with AI21 before relying on them.
What happened to Wordtune?
AI21 stopped developing Wordtune, its writing assistant, in April 2025, according to Israeli press reports. It is no longer part of the company's main business.
Is AI21 Labs part of the AI safety field?
Not as a safety lab. Groups like METR test frontier models for dangerous capabilities; AI21's safety related work is commercial, such as the October 2026 project to make a model resist prompt injection. For the bigger picture, read what AI safety is.
