Jamba is AI21 Labs’ family of foundation models, notable for being one of the first production-scale language models built on a hybrid architecture rather than a pure Transformer. Documented in a March 2024 paper, “Jamba: A Hybrid Transformer-Mamba Language Model,” it interleaves Transformer blocks with Mamba state-space blocks and adds a mixture-of-experts layer. The original release had about 52 billion total parameters with roughly 12 billion active per token, supported a 256,000-token context window, and was engineered to fit on a single 80GB GPU - a combination of long context and small memory footprint that pure Transformers struggle to match.
The hybrid design directly targets the cost of long contexts: Mamba’s state-space layers scale better with sequence length than attention, so Jamba delivers higher throughput on long inputs. AI21 released the weights under a permissive license and has since iterated the line - Jamba 1.6 in March 2025, a small Jamba Reasoning 3B in October 2025, and the Jamba2 family announced in early 2026. AI21’s earlier flagship was the unrelated Jurassic series, including the 178-billion-parameter Jurassic-1 Jumbo of 2021.
Why business readers should care: Jamba is a leading example of moving beyond the standard Transformer to control the cost of long-context processing, which matters for document-heavy enterprise tasks where feeding large amounts of text into the model is the main expense.