Jamba 1.5 Mini

Jamba 1.5 Mini is a hybrid large language model developed by AI21 Labs, utilizing the proprietary Jamba architecture. This architecture combines State Space Model (SSM) layers, specifically Mamba, with Transformer layers and a Mixture-of-Experts (MoE) module. This design is engineered to deliver the high-quality reasoning capabilities of Transformer models while benefiting from the efficiency and high throughput of Mamba-based systems, particularly during long-context processing.

The model features a 256,000-token context window, which is among the largest for open-weight models in its class. It contains 52 billion total parameters, but only 12 billion parameters are active during inference, allowing it to maintain a balance between performance and computational requirements. This efficiency enables the model to handle contexts up to 140,000 tokens on a single 80GB GPU without specialized quantization.

Jamba 1.5 Mini is instruction-tuned and supports a variety of enterprise-ready features, including function calling, structured JSON output, and grounded generation with citation modes. It is a multilingual model, natively supporting languages such as English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, and Hebrew. The model is released under the Jamba Open Model License, permitting research and commercial use.