Jamba by AI21 Labs

Jamba by AI21 Labs is an open-source AI foundation model built on a hybrid Mamba-Transformer (SSM-Transformer) architecture for developers who need long-context performance without prohibitive hardware. The headline capability is handling 256K tokens of context at roughly 2.5x the speed of comparable transformers while fitting on a single 80GB GPU, with weights published openly and results validated on benchmarks. AI21 Labs ships Jamba as a family of sizes — Mini variants around 52B parameters, Large variants around 399B, plus smaller 3B models including a Jamba Reasoning 3B — with quantized FP8 and GGUF formats for reduced resource requirements. The hybrid design blends State Space Models with Transformer layers to keep long-context inference efficient and steerable, which AI21 positions for enterprise reliability. As an open-source long-context LLM, Jamba is aimed at engineers building retrieval, document-analysis, and agentic applications where context length and throughput matter, and it competes with other open foundation models rather than closed hosted APIs alone.

Running an open-source long-context LLM for retrieval, document analysis, and agentic apps that need 256K-token context and high throughput