In short
Aya Expanse is a 32B parameter multilingual model developed by Cohere For AI that bridges the performance gap between English-centric and global languages. It is best suited for global enterprises and researchers needing high-fidelity support for 23 specific languages on mid-range hardware.
Frequently asked
- Which languages does Aya Expanse support?
- The model natively supports 23 specific languages, including Arabic, Chinese, Hindi, Japanese, Korean, and Vietnamese. It focuses on depth of understanding for these languages rather than superficial breadth.
- What are the hardware requirements for running Aya Expanse?
- The model is designed for mid-range hardware deployments, specifically targeting a single or dual GPU setup. However, it requires significant VRAM, approximately 64GB, for full precision inference.
- How is Aya Expanse trained to handle multiple languages?
- It uses a teacher-student framework where a larger 'Aya Expedition' model distills reasoning into the 32B architecture. It is also fine-tuned on the Aya Dataset, a large human-annotated multilingual instruction-following dataset.
- What are the licensing terms for Aya Expanse?
- The model weights are free to download for non-commercial purposes via Hugging Face. Commercial use is restricted by the non-commercial license, though custom pricing is available for high-volume commercial licenses through Cohere.
- What is the context window size for Aya Expanse?
- The model features a large 128k context window, which allows it to process extensive documents such as entire books or codebases.