Transformers have long been considered the foundation of modern generative AI. However, AI21 is introducing a new approach with its Jamba model. Jamba combines transformers with a Structured State Space (SSM) model approach, offering improved performance and accuracy. The recent release of Jamba 1.5 mini and Jamba 1.5 large builds on the initial innovations of Jamba 1.0. Jamba utilizes an SSM approach called Mamba, with the goal of merging the best attributes of transformers and SSM.
The Jamba models, available under an open license, have several new features, including function calling, JSON mode, structured document objects, and citation mode. These additions make the models ideal for crafting agentic AI systems, which require structured data handling and enhanced accountability. JSON mode allows developers to build structured input/output relationships between different parts of a workflow, while citation mode attributes relevant parts of generated content to the documents used. This integration provides more transparency and traceability compared to traditional language models.
It is important to note that citation mode in Jamba 1.5 differs from Retrieval Augmented Generation (RAG). While both approaches ground responses in data to improve accuracy, citation mode in Jamba 1.5 is more tightly integrated with the model itself. In a traditional RAG setup, developers connect the language model to a vector database to access relevant documents, while the model learns to incorporate the retrieved information. However, Jamba 1.5 not only retrieves and incorporates relevant documents but also explicitly cites the sources of information used in its output. This ensures greater transparency and traceability in the model’s reasoning.
AI21 also supports RAG and offers its own end-to-end RAG solution as a managed service. The company has partnerships with major cloud providers like AWS, Google Cloud, and Microsoft Azure, as well as Snowflake, Databricks, and Nvidia. Looking ahead, AI21 plans to advance its models to meet customer needs and continue focusing on enabling agentic AI systems. The company acknowledges the importance of planning and execution in this domain and aims to push the envelope in agentic AI. By constantly innovating and combining different AI approaches, AI21 is shaping the future of AI development.