Cohere releases Aya Vision: 8B/32B multilingual multimodal VLMs

AI Impact Summary

Cohere is releasing Aya Vision, a new family of 8B and 32B parameter vision-language models designed to address multilingual multimodal AI challenges. The models leverage synthetic annotations, data translation, and model merging techniques to achieve superior performance across 23 languages, outperforming models like Llama-3.2 90B Vision and Qwen2.5-VL 72B on benchmarks such as AyaVisionBench and mWildVision. This release represents a significant advancement in multilingual multimodal understanding, particularly for the research community.