Jamba 1.5 MinivsJamba 1.5 Mini (SSM)

Jamba 1.5 Mini is AI21 Labs' efficient model sharing the same hybrid SSM-Transformer architecture as its larger sibling. At 52 billion total parameters with only 12 billion active, it delivers strong performance on long-context tasks while being highly cost-effective. The 256K context window combined with low inference costs makes it particularly attractive for applications processing large volumes of documents.

Jamba is AI21 Labs' pioneering hybrid architecture that combines Mamba state-space model layers with transformer attention layers. This novel approach achieves 3x throughput improvement over pure transformer models of equivalent quality. The Jamba architecture represents a significant step toward more efficient AI models that can process long contexts without the quadratic cost of full attention.