Recent embedding models are initialized from large language model (LLMs), enjoying the massive knowledge encoded in LLMs weights. However, prior works only consider non-reasoning model, and they didn’t study the impact of employing a reasoning model (trained with Reinforcement Learning with Verifiable Reward (RLVR)) as a base model. In this work, we want to answer one question: can a reasoning model lead to a better embedding model? To answer this question, we will explore whether reasoning capabilities can improve the semantic richness and performance of embedding models, by investigating RLVR’s impact on embedding space, and training multiple embedding models to compare the performance difference.