AI21 Jamba

Hybrid SSM-Transformer model with 256K context for long-document processing

Code & DevelopmentFreemium

Jamba is AI21 Labs' hybrid model architecture that combines Mamba State Space Model blocks with standard transformer attention layers, achieving significantly lower memory usage and faster inference for long contexts. Jamba's 256K token context window handles entire books, large codebases, or lengthy legal documents in a single prompt. Available via the AI21 API and in open weights form on Hugging Face. Jamba outperforms similarly-sized pure-transformer models on long-context benchmarks while requiring less GPU memory.

Key Features

• ✓256K context window
• ✓Hybrid SSM-Transformer
• ✓Lower memory usage
• ✓Open weights
• ✓Long document processing
• ✓API + self-host