Jamba Large

Jamba Large is the flagship member of AI21 Labs' Jamba family of foundation models, positioned as the most capable option for complex, enterprise-scale tasks. It is built on a novel hybrid Mamba-Transformer (also described as SSM-Transformer) architecture that interleaves state-space model layers with transformer attention, a design choice AI21 uses to combine the throughput advantages of Mamba with the reasoning quality of attention. The model was introduced as part of the Jamba 1.6 family, with the larger variant named Jamba Large 1.6 in AI21's launch announcement, and the same family lineage has continued through later iterations referenced in third-party reporting.

A defining trait of Jamba Large is its 256,000-token context window, which AI21 highlights as the longest available among open-weight models at the time of announcement. That capacity, paired with the hybrid architecture, makes the model well suited to long-context enterprise workloads such as retrieval-augmented generation and grounded question answering over large document collections. Because the weights are openly published and downloadable from Hugging Face, organizations in regulated industries like finance and healthcare can run inference on-premises or inside their own virtual private cloud, keeping sensitive data fully under their own control while still accessing a model AI21 describes as competitive with leading closed systems on quality benchmarks such as Arena Hard and long-context evaluations.