AI21 Labs, the Israeli lab behind the Jamba models

An Israeli company focused on the enterprise

AI21 Labs is an AI company associated with Tel Aviv, co-founded by Ori Goshen, its CEO, and Yoav Shoham, a professor and the company’s chairman. Its official pages do not state a founding year; secondary sources date it to 2017. AI21 describes itself as a pioneer of “enterprise AI systems and foundation models,” with a focus on trustworthy AI that boosts productivity. Its investors include Google, NVIDIA, Intel Capital and Samsung.

Jamba, a Transformer and Mamba hybrid

The Jamba family is AI21’s technical signature. Most large language models rely solely on the Transformer architecture, whose cost climbs quickly as text gets longer. Jamba mixes Transformer layers with Mamba layers, a state space model (SSM) that is more efficient on long sequences. The largest versions add a mixture of experts. The goal is a long context, suited to technical manuals or bulky case files, without runaway memory needs.

As of October 22, 2026, the models published on Hugging Face include:

• Jamba2 (Mini and 3B), released in January 2026 under the Apache 2.0 license;
• Jamba Reasoning 3B, a small reasoning model;
• the Jamba 1.5, 1.6 and 1.7 generations (Mini and Large), under the Jamba Open Model License.

A strategy that has shifted

In 2026, AI21’s headline offer is no longer just a model catalog. The company sells a post-training service it calls “lab-as-a-service”: its researchers build evaluations from the client’s real tasks, train a specialized model, optimize it for production and improve it with usage data. It also develops Maestro, an AI agent system geared toward deep research. In practice, AI21 is betting that many companies will gain more from a specialized model than from a bigger general-purpose one.

How to access the models

Jamba weights can be downloaded from Hugging Face. The API and guides live in AI21’s documentation. To compare Jamba with other open models, see our page on open-source models.