Cohere for AI (C4AI) has made a groundbreaking announcement today with the release of Aya 23, a new family of state-of-the-art multilingual language models. These models, available in 8B and 35B parameter variants, aim to expand language modeling capabilities to nearly half of the world’s population.
Aya 23 builds on the success of the original Aya 101 model and serves 23 languages, including Arabic, Chinese, English, French, German, Japanese, Russian, Spanish, and more. By open sourcing the weights of Aya 23, C4AI is allowing third-party researchers to fine-tune the models to fit their specific needs, marking a significant step forward in the field of multilingual language modeling.
According to C4AI, Aya 23 outperforms not only its predecessor Aya 101 but also other open models like Google’s Gemma and Mistral’s various open source models. The models have shown higher-quality responses across the languages they cover, breaking language barriers and improving performance on discriminative and generative tasks.
The release of Aya 23 represents a shift towards balancing breadth and depth in language modeling. By focusing on fewer languages with more capacity, the models have shown significant improvements in performance compared to Aya 101 and other widely used models.
Researchers at C4AI have reported that Aya 23 achieves a 6.6x increase in multilingual mathematical reasoning compared to Aya 101 and consistently outperforms other models in various benchmarks. The open weights for both the 8B and 35B models are now available on Hugging Face under a Creative Commons license, allowing researchers and practitioners to advance multilingual models and applications.
Overall, the release of Aya 23 represents a major milestone in the field of multilingual language modeling, offering new possibilities for researchers and developers to create more inclusive and high-performing AI models.