The Jamba 1.5 launch in August 2024 marked a significant step in the evolution of large language models by AI21 Labs, an Israeli company specializing in natural language processing. The release introduced two models, Jamba 1.5 Large and Jamba 1.5 Mini, designed to offer high performance with improved efficiency and long-context handling. These models are built on a hybrid architecture that combines the strengths of transformer and Mamba state space models, using a mixture-of-experts approach to balance computational cost and output quality.
Jamba 1.5 models support a context window of up to 256,000 tokens, enabling them to process and reason over very long documents, such as entire books or extensive codebases. This capability, paired with a focus on efficiency, positions Jamba 1.5 as a competitive option among open-weight models, particularly for enterprise applications that require processing large volumes of text.
Background and Development
AI21 Labs was founded in November 2017 by Yoav Shoham, Ori Goshen, and Amnon Shashua in Tel Aviv, Israel. The company initially gained attention with Wordtune, an AI-based writing assistant launched in October 2020, which Google later named one of its favorite extensions of 2021. In August 2021, AI21 Labs introduced AI21 Studio and the Jurassic-1 language model, which featured a token vocabulary of over 250,000.
The company's trajectory included several funding rounds: a $9.5 million seed round in January 2019, a $20 million investment from Walden Catalyst in November 2021, a $25 million Series A led by Pitango First, a $64 million Series B in July 2022, and a $155 million Series C in August 2023, with participation from Google and Nvidia. These investments supported the development of increasingly sophisticated models, including the release of Jurassic-2 in March 2023.
The direct predecessor to Jamba 1.5 was the original Jamba model, released on March 29, 2024. Jamba was an open-weights model built on a hybrid Mamba-SSM transformer architecture with a mixture-of-experts design, supporting context lengths up to 256,000 tokens. The Jamba 1.5 family refined this approach, improving efficiency and performance.
Technical Architecture
The Jamba 1.5 models employ a hybrid architecture that integrates multi-head attention from transformers with Mamba state space models. This combination aims to capture the strengths of both: the ability of transformers to model complex dependencies and the efficiency of state space models in handling long sequences. The mixture-of-experts (MoE) mechanism activates only a subset of the model's parameters for each token, reducing computational cost while maintaining high capacity.
This design allows Jamba 1.5 to process long contexts without the quadratic scaling typical of pure transformer models, making it more efficient for tasks like document summarization, question answering over lengthy texts, and code analysis. The models were trained on a diverse dataset, and their open-weights nature allows researchers and developers to fine-tune them for specific applications.
Release and Availability
Jamba 1.5 was released in August 2024, with two variants: Jamba 1.5 Large and Jamba 1.5 Mini. The models were made available under an open license, allowing broad use and modification. In September 2024, AI21 Labs partnered with Amazon Web Services (AWS) to offer the Jamba 1.5 family on Amazon Bedrock, a managed service that enables integration of AI models into enterprise applications. This partnership facilitated deployment for businesses seeking to leverage Jamba's long-context capabilities.
The release positioned Jamba 1.5 as a competitor to other open-weight models from organizations like OpenAI, Anthropic, and Google DeepMind, though those companies primarily offer proprietary models. Jamba 1.5's emphasis on efficiency and long-context handling appealed to developers who needed to process large datasets without incurring high computational costs.
Performance and Benchmarks
AI21 Labs reported that Jamba 1.5 models achieved competitive results on various benchmarks, including natural language understanding, reasoning, and code generation tasks. The models were particularly noted for their performance on long-context tasks, where they could maintain accuracy over sequences of up to 256,000 tokens. This was a key differentiator, as many models at the time struggled with such extended contexts.
The efficiency gains from the hybrid architecture allowed Jamba 1.5 to deliver these results with lower inference costs compared to models of similar capability. This made it attractive for organizations with limited computational resources.
Enterprise Adoption and Partnerships
The partnership with AWS in September 2024 was a major step toward enterprise adoption. By making Jamba 1.5 available on Amazon Bedrock, AI21 Labs enabled companies to integrate the models into their workflows with the security and scalability of AWS infrastructure. This move was part of a broader trend of AI companies partnering with cloud providers to reach a wider audience.
In March 2025, AI21 Labs released Jamba 1.6 for private enterprise deployment, claiming it outperformed other open models across multiple benchmarks. The same month, the company launched Maestro, an AI planning and orchestration system aimed at improving the accuracy of models like GPT-4o and Claude 3.5 Sonnet in complex tasks. These developments indicated AI21 Labs' continued focus on enterprise solutions.
Impact and Reception
The Jamba 1.5 launch was well-received in the AI community, particularly for its innovative hybrid architecture and long-context capabilities. It contributed to the growing interest in state space models as an alternative to pure transformer architectures, which had dominated the field since the introduction of the transformer in 2017. Researchers and practitioners appreciated the open-weights approach, which allowed for experimentation and customization.
The models also sparked discussions about efficiency in AI, as the mixture-of-experts design reduced the environmental and financial costs of running large models. This aligned with broader efforts in the industry to make AI more sustainable.
Future Directions
Following Jamba 1.5, AI21 Labs continued to iterate on its model family, releasing Jamba 1.6 in March 2025 with improved performance for enterprise use. The company also expanded its product ecosystem with Maestro, which aimed to enhance the reliability of AI systems in complex, multi-step tasks. These efforts suggested a strategy focused on practical, deployable AI solutions rather than just pushing benchmark scores.
The hybrid architecture pioneered in Jamba models may influence future developments in machine learning, as researchers explore ways to combine different model types to achieve better efficiency and capability. The success of Jamba 1.5 demonstrated that state space models can be effectively integrated with transformers, opening new avenues for research.
See Also
- Artificial intelligence
- Generative AI
- AI21 Labs
- Mamba state space models (if available)
References
- AI21 Labs official documentation and press releases.
- Industry reports on the Jamba 1.5 launch.
- Academic papers on hybrid transformer-state space models.