Jamba

Jamba is AI21's language model family for teams building document-based assistants and business applications on their own servers. The Large model targets lengthy context, but its hardware requirements put it in server territory: quantized deployment uses eight GPUs with 80GB of memory each.

• Long-document context: A 256K-token context window supports work such as investment research, due diligence and procurement response review.
• Context-grounded answers: The model is designed to base responses on supplied material and follow instructions about response style, useful for document assistants and customer-facing chatbots.
• Multilingual generation: Work in English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic or Hebrew for tasks such as product descriptions, reports and tutoring.
• Model customization: LoRA, qLoRA and full fine-tuning provide ways to adapt the model to your application.

Jamba combines Mamba state-space models with Transformer attention for efficient long-context processing. Private deployment runs inference on your infrastructure through vLLM or Hugging Face Transformers. With ExpertsInt8 quantization and the default vLLM configuration, the Large model supports prompts up to 220K tokens on eight 80GB GPUs, below its full context window.

The Jamba Open Model License permits research and commercial use under its terms. Access to model files on Hugging Face requires an account, acceptance of the access conditions and sharing your contact information.