Jamba Large 1.6

Jamba Large 1.6 is AI21 Labs’ flagship language model, built on a hybrid architecture that combines transformer attention layers with Mamba state-space model layers. AI21 released it in March 2025 as part of the Jamba 1.6 family, alongside a smaller Jamba Mini 1.6. The model has 398 billion total parameters, with only 94 billion active at inference time thanks to a mixture-of-experts design, which keeps compute costs down while supporting a 256,000-token context window. AI21 markets it toward enterprise customers who need to process long documents, contracts, or codebases without sending data to a third party, since the model is available for self-hosted deployment as well as through an API. In AI21’s own benchmarking, Jamba Large 1.6 beat open models from Meta, Mistral, and Cohere on a range of tasks, and it runs notably faster than similarly sized dense transformers because of the Mamba layers’ more efficient handling of long sequences.