Cohere's Aya Vision beats models twice its size in 23-language image tests

Cohere For AI just dropped a pair of open-weight vision-language models that punch well above their weight class — and do it across 23 languages. The Aya Vision 8B and 32B models, released today, don’t just match competitors in their size bracket; in head-to-head tests, the 32B version beats models more than twice its size, including Llama-3.2 90B Vision and Qwen2.5-VL 72B, with win rates between 50% and 64% on the new AyaVisionBench benchmark.

The secret sauce is a training pipeline built on synthetic data and a clever multilingual lift. Cohere’s team started with high-quality English annotations, then translated and rephrased them into 23 languages to avoid the stilted, error-ridden text you often get with straight machine translation. That rephrasing step matters more than it sounds — it preserves fluent answer quality while expanding coverage into languages where real-world multimodal datasets barely exist. The 8B model, when fine-tuned only on standard academic datasets, managed a 40.9% win rate against Pangea 7B. After this data enhancement, it reportedly hits up to 79%.

Architecture-wise, Aya Vision uses SigLIP2 as its vision encoder and employs Pixel Shuffle to compress image tokens by 4x, tackling the latency problem that plagues high-resolution processing. For the language backbone, the 8B model starts from Cohere Command R7B — post-trained with the Aya Expanse multilingual recipe — while the 32B builds directly on Aya Expanse 32B. Both go through a two-stage training: first aligning vision to language by training only the connector, then full supervised fine-tuning on multimodal tasks.

What stands out to me is the decision to release both AyaVisionBench and a multilingual version of Wild Vision Bench alongside the models. That transparency makes the performance claims testable. It also shifts the multilingual multimodal race away from proprietary API pricing toward open-weight efficiency. If a 32B model can genuinely hold its own against 90B proprietary alternatives, the cost-to-performance equation changes for anyone building in languages beyond English.