AI21 Jamba: Model Facts and Web Access for Site Owners
Table of Contents
Jamba is a family of open language models from AI21 Labs. A model has no web access of its own: it reads only the text that an application sends to it. AI21 does not document a crawler or a user agent, so a site owner has no Jamba rule to add.
What Jamba is
AI21’s Jamba documentation describes a Mamba-Transformer architecture and a context window of 256K tokens. A context window is the amount of text that the model can read in one request. The models take text as input and give text as output.
| Model | Size | Version |
|---|---|---|
| Jamba Large | 398B parameters (94B active) | 1.7 |
| Jamba Mini | 52B parameters (12B active) | 2 |
| Jamba 3B | 3B parameters | 2 |
The documentation gives a knowledge cutoff date of August 22, 2024. Companies can run the models in a private cloud or on their own servers, or call them through the AI21 API.
Web access
Jamba itself does not search the web or open pages. It answers from its training and from the text in the request. If an application gives Jamba web content, the application fetches that content, not the model.
AI21 has a separate product, Maestro, with a web search tool in AI21 Studio. A user can enter the websites that the tool searches. AI21 does not document a user agent, an IP list or robots.txt behavior for this tool.
What a site owner can do
AI21 publishes no crawler token, so there is no Jamba group to add to robots.txt. robots.txt is a text file at the site root with rules for crawlers. If a self-hosted Jamba application fetches your pages, it uses the user agent that its developer set, not one from AI21.