AI21 Labs

AI21 Labs sells its Jamba large language models through a usage-based API, with open weights for private deployment, for enterprise document and agent work.

Jamba models take long contracts, filings and manuals in a single request, so an answer can draw on the full document rather than scattered fragments. AI21 Labs Ltd. trains the Jamba family and serves it through AI21 Studio, a developer platform with an API, an SDK and a web console. The models are aimed at enterprise work where answers have to stay grounded in a company's own material. Jamba Large is the long-context model for the hardest jobs, while Jamba Mini is the lightweight model for a wide range of everyday tasks. Both take text prompts and support function calling, returning structured calls to a developer's own tools. The weights are published on Hugging Face under the Jamba Open Model License, and a fine-tuning guide covers full, LoRA and QLoRA training on downloaded weights. The File Library stores uploaded documents, including files synced from Google Drive, and Maestro runs have an endpoint of their own. Around the models, the Intelligent Gateway gives live agent traffic one endpoint for routing and compaction, and the Dedicated Model service trains a compact open model for one customer's defined workloads. Enterprise engineering teams building retrieval and agent systems over large document collections are the core audience, financial analysis included. Teams with strict data rules can download the weights and run them in their own environment instead of calling a hosted service. Buying is by usage: a developer starts on trial credits without a card, then pays per input and output token for each model, with no limit on seats. Larger buyers move to the Custom Plan, a sales contract that adds volume discounts, higher rate limits, private cloud hosting and a dedicated account manager. Grounded question answering over long documents is the central job, along with classification and retrieval-augmented generation over a knowledge base. Companies that already buy AI through Amazon Bedrock or Google Vertex AI can reach Jamba from inside those cloud accounts. Teams that self-host the open weights deploy them with vLLM on their own infrastructure. A dated deprecation table lists when each model version retires, so a team can plan a migration before an endpoint shuts down. Jamba fits document-heavy companies that want long-context models they can either rent by the token or run themselves.