Jamba
Hybrid open-weights language model with long context
Jamba is a large language model released by AI21 Labs in March 2024. It uses a hybrid architecture combining Mamba state-space models and transformers with mixture of experts. The model supports context windows up to 256,000 tokens and offers open weights for developers.
- Released by AI21 Labs in March 2024
- Uses hybrid Mamba and transformer architecture
- Supports context windows up to 256,000 tokens
- Available via Amazon Bedrock since 2024
- Open weights model for developer access
What it is
Jamba is a large language model developed by AI21 Labs, an Israeli artificial intelligence company based in Tel Aviv. First released on 29 March 2024, it represents a shift in design by combining different architectural approaches. The model is built for general-purpose natural language tasks and supports extremely long input sequences. It allows developers to access its underlying weights openly, facilitating custom integration and research. Later versions, including Jamba 1.5 and 1.6, expanded its availability through cloud platforms and private enterprise options.
How it came about
AI21 Labs, founded in 2017, previously launched the Jurassic series of models. Jurassic-1 arrived in August 2021, followed by Jurassic-2 in March 2023. Jamba was announced on 29 March 2024 as a new direction for the company. In September 2024, the Jamba 1.5 family became available on Amazon Bedrock through a partnership with AWS. This move enabled enterprises to integrate the model into their applications. By March 2025, AI21 Labs released Jamba 1.6, targeting private enterprise deployment with claims of superior performance against other open models.
Inside the model
The model employs a hybrid architecture that merges Mamba state-space models with standard transformer layers. State-space models are efficient for processing long sequences, while transformers excel at capturing complex relationships in data. Jamba also uses a mixture of experts technique, where only a subset of the network activates for each token. This design aims to balance computational efficiency with high performance. The system supports context lengths of up to 256,000 tokens, allowing it to process vast amounts of text in a single pass without losing earlier information.
What it can do
Jamba is designed to handle complex natural language processing tasks with improved efficiency. The March 2025 release of Jamba 1.6 claimed to outperform other open models across multiple benchmarks. Its long context window enables it to analyse extensive documents or codebases in one go. The model supports multilingual inputs, reflecting the capabilities of its predecessor, Jurassic-2. By leveraging the hybrid design, it aims to deliver faster response times and better instruction following than previous dense architectures.
How to use it
The initial release featured open weights, allowing researchers and developers to download and modify the model. In September 2024, AI21 Labs partnered with Amazon Web Services to host the Jamba 1.5 family on Amazon Bedrock. This provided an API-based route for enterprise users who prefer managed services over self-hosting. Jamba 1.6, released in March 2025, focused on private enterprise deployment, offering organisations more control over their data and infrastructure while utilising the model's capabilities.
Why it matters
Jamba introduced a notable hybrid architecture to the mainstream market, blending state-space models with transformers. This approach influenced discussions on how to scale models efficiently without relying solely on dense transformer designs. Its availability on major cloud platforms like AWS helped broaden its adoption in enterprise settings. The open-weights release encouraged community experimentation with the Mamba architecture, potentially accelerating research into more efficient large language model structures.
Questions people ask about Jamba
What makes Jamba's architecture different?
Jamba uses a hybrid design that combines Mamba state-space models with transformer layers. It also employs a mixture of experts, activating only part of the network for each token. This structure aims to improve efficiency and handle long contexts better than traditional dense transformers.
How long is the context window for Jamba?
Jamba supports context lengths of up to 256,000 tokens. This allows the model to process very large documents or extensive code files in a single interaction, maintaining awareness of information from the beginning of the input sequence.
Is Jamba available for commercial use?
Yes. Jamba 1.5 is available on Amazon Bedrock for enterprise integration via API. Jamba 1.6, released in March 2025, is designed for private enterprise deployment. The model also offers open weights, allowing developers to self-host and modify it.
Good reading and sources
- A.I. Bots Can’t Report This Column. But They Can Improve It. (Published 2023)The New York Times, Article
- AI21 Labs nabs $64M to ramp up AI-powered language servicesTechCrunch, Article
- Watch out, GPT-3, here comes AI21's 'Jurassic' language modelZDNET, Article
- AI21 Labs writing tool will not do all the hard work for youThe Register, Article
The facts on this page come from the Wikipedia article "AI21 Labs", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.