AI21 Introduces Jamba 1.6, Raising the Bar for Accuracy and Speed in Open Models

Jamba 1.6 delivers enterprise-grade AI performance, surpassing open models from Mistral, Meta, and Cohere across multiple benchmarks—without compromising speed or data control.

TEL AVIV, Israel, March 6, 2025 /PRNewswire/ -- AI21 has launched Jamba 1.6, the most advanced open model for private enterprise deployment, designed to meet real-world business needs without sacrificing security or performance. Built on AI21's hybrid SSM-Transformer architecture, Jamba 1.6 delivers industry-leading results, outperforming leading models across key enterprise use cases while providing the best accuracy, speed, and security in the market.

Jamba 1.6 outperforms its open model competitors from Mistral, Meta, and Cohere across multiple benchmarks, showcasing superior general quality (Arena Hard) and best-in-class performance on retrieval-augmented generation (RAG) and long-context question answering (QA). Critically, Jamba 1.6 achieves these results without compromising on speed, as demonstrated in scatter plot benchmarks.

One of the biggest challenges enterprises face when adopting AI is maintaining full control over their data. As an open model, Jamba 1.6 can operate entirely within a company's private environment (self-hosted), with flexible deployment options including VPC and on-premise installations.

The new model improves data classification by 26 percentage points over AI21's previous model - Jamba 1.5, enabling more accurate structuring and automation of large datasets. Its ability to process large volumes of unstructured data with high accuracy makes it ideal for summarization and document analysis.

Jamba 1.6 delivers cited responses with over 90% consistency across long contexts, ensuring reliability in grounded question answering. It also integrates seamlessly with enterprise knowledge bases through retrieval-augmented generation (RAG), providing precise, context-aware insights.

"Jamba 1.6 delivers unmatched speed and performance, setting a new benchmark for enterprise AI," said Or Dagan, Chief Product & Strategy Officer of AI21. "With this release, we're proving that enterprises can achieve exceptional AI capabilities without compromising efficiency,  security, and data privacy."

Transformers are the backbone of today's most powerful AI models, excelling at weighing the relevance of different pieces of input data when generating responses. However, they are computationally expensive, especially when processing long sequences of data. State Space Models (SSMs), on the other hand, offer a more efficient approach by condensing the context into a fixed state. But while pure SSMs show promising results, they still demonstrate quality gaps compared to Transformers.

By combining the strengths of both architectures, Jamba 1.6 leverages the precision and reasoning capabilities of transformers with the efficiency and scalability of SSMs. This hybrid approach allows for better handling of long-context tasks while maintaining high performance, making it a powerful solution for enterprise AI applications.

About AI21

AI21 is a pioneer in Foundation Models and AI Systems designed for enterprises. AI21's mission is to create trustworthy artificial intelligence that powers humanity towards superproductivity. Founded in 2017 by AI visionaries Prof. Amnon Shashua, Prof. Yoav Shoham, and Ori Goshen, AI21 has secured $336 million in funding from industry leaders, including NVIDIA, Google, and Intel, reinforcing its commitment to advancing AI innovation.

SOURCE AI21 Labs