A "modern" reranker, supporting 8K context, multiple languages, code, json etc. It is also trained with GRPO, which is a first for this kind of model.
