AI21 Lab has recently released upgraded versions of the Jamba model - Jamba 1.5 Mini and Jamba 1.5 Large. These two new versions cleverly combine the advantages of the Transformer architecture and the Structured State Space (SSM) model, aiming to improve performance and accuracy. The core of the Jamba series lies in its unique Joint Attention and Mamba (Jamba) architecture, which is achieved by introducing a method called Mamba and combining it with Transformer technology. Since the release of Jamba 1.0 in March this year, AI21 Lab has received extensive feedback from the industry. As the first attempt to apply Mamba technology in a large-scale production environment, the Jamba architecture has sparked heated discussions within the industry about the future direction of language models. The newly launched Jamba 1.5 series adds several new features to the existing foundation, including support for function calls, JSON schema handling, structured document object manipulation, and reference patterns, aiming to assist in building more complex "agent-based" artificial intelligence systems. Both Jamba 1.5 Mini and Large are equipped with a large context window of 256K and adopt an advanced Mixture-of-Experts (MoE) architecture. Specifically, the Mini version has a total of 52 billion parameters, with 12 billion active parameters, while the Large version is more powerful, with a total of 398 billion parameters, of which 94 billion are active. These two models are provided under an open license, and AI21 Lab also offers commercial support and services to users. In addition, AI21 has established partnerships with several well-known cloud service providers (such as AWS, Google Cloud, Microsoft Azure) and data processing platforms (such as Snowflake, Databricks, Nvidia). The main new features introduced in Jamba 1.5 include: - JSON schema: Enhances the model's ability to process structured data, making it easier for developers to build and manage input/output relationships in complex workflows. - Reference patterns: Improves the traceability and transparency of the system, ensuring that the generated content can accurately annotate its information source. - Document API: Optimizes the context management mechanism, allowing the model to more effectively reference relevant documents when generating text. - Function call capability: Empowers the model to perform specific tasks, such as database queries or computational operations. Or Dagan, Vice President of Product at AI21 Lab, pointed out that the addition of JSON schema helps developers build application workflows more efficiently. Meanwhile, the introduction of reference patterns is aimed at ensuring that the content generated by the model can be clearly annotated with its information source, enhancing the credibility and traceability of the content. It is worth noting that although both reference patterns and Retrieval Augmented Generation (RAG) techniques involve retrieving information from external data sources to improve the accuracy of generated content, there are significant differences in their implementation. RAG typically relies on vector databases to retrieve relevant documents and the model learns how to integrate this information. On the other hand, Jamba 1.5's reference patterns are more tightly integrated within the model itself, not only achieving information retrieval and integration but also providing clear annotations of the information source, thus offering higher transparency and traceability. In addition, AI21 Lab also provides an end-to-end RAG solution as a managed service, which includes document retrieval, indexing, and other functionalities.