AI21 Jamba

Jamba is AI21’s open model family built on a hybrid state-space-model and Transformer architecture. It is designed for memory-efficient long-context processing, reliable instruction following, grounding, and high-throughput enterprise workloads. Jamba2 offers compact dense and mixture-of-experts variants for question answering, retrieval workflows, and production agents without requiring heavy reasoning-token overhead. Weights are available for download, and hosted access is offered through AI21 Studio.