A large-scale mixture-of-experts vision-language model that activates only 22B parameters at a time despite its 235B total size, keeping inference costs manageable. It handles both text and images with the thoroughness you'd expect from a model of this scale — detailed visual analysis, complex reasoning across modalities, and nuanced instruction following. The MoE architecture means it punches above its active-parameter weight, though deploying it still demands serious infrastructure.