Beyond Transformers: How AI21’s Jamba 1.5 is Redefining Generative AI

AI21 Labs is making waves again with the release of Jamba 1.5, an evolution of its groundbreaking model that combines transformers with a Structured State Space Model (SSM) approach called Mamba. With two new versions, Jamba 1.5 Mini and Jamba 1.5 Large, AI21 is redefining what’s possible for large language models (LLMs) by merging powerful architectures for improved performance, accuracy, and usability.

According to Or Dagan, VP of Product at AI21, Jamba remains one of the only production-scale models built on Mamba. “It’s a novel architecture that started debates about the future of LLMs,” he shared. The hybrid design challenges the dominance of transformers by integrating complementary capabilities from SSMs, opening the door to more efficient, adaptable AI systems.

Key Innovations in Jamba 1.5

With the Jamba 1.5 series, AI21 introduces several features aimed at accelerating the development of agentic AI systems:

• JSON Mode

  : Enables seamless structured data handling, making it easier for developers to build workflows with well-defined input/output relationships.

• Citation Mode

  : Goes beyond traditional Retrieval Augmented Generation (RAG) by tightly integrating document retrieval and attribution within the model itself. This ensures transparency, accountability, and traceability.

• Document API

  : Enhances context management by allowing the model to process and attribute content from structured documents.

• Function Calling

  : Adds versatility for developers building interactive and dynamic AI-driven systems.

Jamba 1.5 models also feature large context windows (256K) and leverage Mixture-of-Experts (MoE) techniques. Jamba 1.5 Mini boasts 52 billion total parameters, with 12 billion active at a time, while Jamba 1.5 Large delivers an impressive 398 billion total and 94 billion active parameters.

Redefining RAG with Citation Mode

One of Jamba 1.5’s standout features is its citation mode, which provides a more integrated and transparent alternative to traditional Retrieval Augmented Generation (RAG). While RAG relies on external vector databases for document retrieval, Jamba 1.5’s citation mode incorporates this functionality directly into the model. This integration allows Jamba 1.5 to retrieve relevant data, generate responses, and explicitly cite sources in its output—all within the same system.

This capability is a game-changer for developers looking to build AI solutions that require a high level of accountability, such as in research, legal, or financial applications. Dagan emphasized that the approach simplifies workflows while ensuring users can trust the model’s reasoning.

Open Licensing and Enterprise Support

AI21 is offering both Jamba 1.5 Mini and Large under an open license, making the models accessible to a wide range of developers and organizations. To ensure success at scale, the company also provides commercial support and services. Partnerships with cloud giants like AWS, Google Cloud, Microsoft Azure, and Nvidia underscore AI21’s commitment to seamless integration for enterprise users.

A Focus on Agentic AI

Looking ahead, AI21 is doubling down on its vision for agentic AI—systems capable of planning and executing complex tasks autonomously. With the innovations in Jamba 1.5, the company is laying the groundwork for a future where AI systems are not only smarter but also more accountable and adaptable to real-world needs.

As AI21 continues to push boundaries, one thing is clear: Jamba is more than just an LLM—it’s a glimpse into the future of AI architecture.