The Jamba 1.5 Open Model Family: The Most Powerful and Efficient Long Context Models

AI21 Labs has debuted the Jamba 1.5 family of open models, Jamba 1.5 Mini and Jamba 1.5 Large, built on a novel hybrid SSM-Transformer architecture. This design combines the quality of Transformers with the efficiency of Mamba, enabling a 256K context window, faster inference speeds, and a lower memory footprint. The models outperform competitors in their size class on benchmarks like Arena Hard and RULER, demonstrating strong performance across their entire context length. To aid deployment, AI21 also developed ExpertsInt8, a new quantization technique tailored for Mixture-of-Experts (MoE) models. The models are available under an open license on Hugging Face and various cloud platforms, supporting enterprise features like function calling and multilingual capabilities.