Cohere for AI Unveils Aya Expanse Models to Close Global Language Gap
Cohere for AI introduces Aya Expanse, an open-weights multilingual model family designed to bring state-of-the-art AI capabilities to underrepresented languages. The 32B model outperforms significantly larger competitors on multilingual understanding benchmarks.
Cohere for AI releases Aya Expanse, a new open-weights family of multilingual language models designed to bridge the persistent language gap in artificial intelligence. Available in 8B and 32B sizes, these models aim to provide equitable AI access to communities speaking low-resource languages that traditional natural language processing tools often ignore. By offering open weights, Cohere for AI allows researchers and developers worldwide to freely access and build upon this technology.
The Aya Expanse models utilize an advanced transformer architecture to handle a wide range of natural language tasks, including text generation, translation, and summarization. To ensure equitable performance across linguistic contexts, the system trains on diverse datasets that incorporate low-resource languages such as Swahili, Bengali, and Welsh. This targeted approach helps capture linguistic nuances and semantic richness that monolingual or English-centric models typically miss.
In rigorous evaluations, the Aya Expanse-32B model demonstrates exceptional capabilities by outperforming much larger models like Gemma 2 27B, Mistral 8x22B, and Llama 3.1 70B. Most notably, the 32B model achieves a 25 percent higher average accuracy across low-resource language benchmarks compared to its competitors, despite being less than half the size of some rivals. This efficiency makes Aya Expanse a highly practical solution for deploying inclusive AI systems across various environments.