Multimodal Models Are Only as Good as Their Weakest Encoder

Multimodal models aren't one model with uniform capability across modalities. They're a pipeline, and production quality is set by the weakest encoder