** "Introducing mcdse-2b-v1: A Multilingual Model for Visual Retrieval"

The development team at marplex has announced the release of mcdse-2b-v1, a visually multilingual model trained on the colpali train set. The new model demonstrates improved performance in understanding broader query distributions and mitigating overoptimizations. Preliminary tests show that mcdse-2b-v1 outperforms its predecessor by 5% on certain queries.

The team notes that training on more multilingual data will likely increase the model's average score and improve performances on ShiftProject. The new model also supports Qwen2VL for embedding tasks, enabling faster inference compared to HuggingFace Transformers.

**

Source: https://dev.to/marplex/visually-multilingual-introducing-mcdse-2b-41gj