The paper introduces a novel framework for transferable Boltzmann Generators (TBGs), which are designed to generate samples from equilibrium distributions of molecular systems without retraining for new systems. The authors extend the work on flow matching for training Boltzmann Generators (BGs) to enable transferability across chemical space. The proposed TBGs are based on continuous normalizing flows (CNFs) and are trained using flow matching, a simulation-free training method. The key contributions include: 1. **Introduction of Transferable Boltzmann Generators**: The first framework for TBGs that can generate unbiased samples from unseen Boltzmann distributions without retraining. 2. **General Framework**: A comprehensive framework for training and sampling with TBGs based on CNFs, including post-processing of generated samples. 3. **Ablation Studies**: Demonstrates that even small training sets can be sufficient for training TBGs, and highlights the importance of including topology information in the architecture for efficient generalization. The authors evaluate the proposed TBGs on dipeptides, showing that they can generate unbiased samples from unseen dipeptides and accurately sample physical properties such as free energy differences between metastable states. The TBGs also demonstrate data efficiency, with even small training trajectories being effective. The paper discusses limitations and future work, including the need to scale to larger systems and explore alternative architectures for the vector field.The paper introduces a novel framework for transferable Boltzmann Generators (TBGs), which are designed to generate samples from equilibrium distributions of molecular systems without retraining for new systems. The authors extend the work on flow matching for training Boltzmann Generators (BGs) to enable transferability across chemical space. The proposed TBGs are based on continuous normalizing flows (CNFs) and are trained using flow matching, a simulation-free training method. The key contributions include: 1. **Introduction of Transferable Boltzmann Generators**: The first framework for TBGs that can generate unbiased samples from unseen Boltzmann distributions without retraining. 2. **General Framework**: A comprehensive framework for training and sampling with TBGs based on CNFs, including post-processing of generated samples. 3. **Ablation Studies**: Demonstrates that even small training sets can be sufficient for training TBGs, and highlights the importance of including topology information in the architecture for efficient generalization. The authors evaluate the proposed TBGs on dipeptides, showing that they can generate unbiased samples from unseen dipeptides and accurately sample physical properties such as free energy differences between metastable states. The TBGs also demonstrate data efficiency, with even small training trajectories being effective. The paper discusses limitations and future work, including the need to scale to larger systems and explore alternative architectures for the vector field.