AI21Labs recently released its latest large language model, the Jamba 1.6 series, touted as the most powerful and efficient long-text processing model currently available. Compared to traditional Transformer models, Jamba boasts significantly faster and higher-quality long-context processing, achieving 2.5 times the inference speed of comparable models, marking a significant technological leap.
The Jamba 1.6 series includes Jamba Mini (1.2 billion parameters) and Jamba Large (9.4 billion parameters), specifically optimized for commercial applications. It features function calling, structured output (like JSON), and reality-grounded generation capabilities. These models are applicable across a wide range of uses, from enterprise-grade intelligent assistants to academic research.
This model is released under the Jamba Open Model License, allowing for both research and commercial use under the terms of the license. Furthermore, the Jamba 1.6 series' knowledge cutoff date is March 5, 2024. It supports multiple languages including English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, and Hebrew, ensuring global accessibility.
In performance evaluations, Jamba Large 1.6 excelled across multiple benchmarks. It outperformed competitors on standard tests like Arena Hard, CRAG, and FinanceBench, demonstrating superior language understanding and generation capabilities. Its performance is particularly noteworthy in handling long-form text.
For efficient inference, users need to install the relevant Python libraries and require a CUDA-enabled device to run the model. The model can be run using the vLLM or transformers framework. Supported by large-scale GPUs, Jamba Large 1.6 can process contexts up to 256K tokens, a feat previously unattainable.
Model: https://huggingface.co/ai21labs/AI21-Jamba-Large-1.6
Key Highlights:
🌟 Jamba 1.6 offers faster and higher-quality long-text processing and supports multiple languages.
🚀 The open-source license allows for research and commercial use, fostering technology sharing.
💡 It outperforms competitors in multiple benchmark tests.
