A Deepdive into Aya Vision: Advancing the Frontier of Multilingual Multimodality
Hugging Face Blog
Cohere For AI released Aya Vision, a family of open-weight vision-language models in 8B and 32B parameter sizes that support image understanding across 23 languages. The 32B model outperforms models 2x its size like Llama-3.2 90B Vision by 50-64% on their AyaVisionBench benchmark, while the 8B model achieves up to 79% win-rates against comparable-sized competitors. Researchers and developers can now access these models and two new multilingual vision-language benchmarks on Hugging Face for building applications like the WhatsApp integration already available.
