This paper introduces a framework for transferable Boltzmann Generators (TBGs) that can generate samples from unseen Boltzmann distributions without retraining. TBGs are based on continuous normalizing flows (CNFs) and are designed to be transferable across chemical space. The framework allows for efficient sampling from the target distribution and reweighting to the unbiased target Boltzmann distribution. The transferability of the framework is evaluated on dipeptides, where it shows efficient generalization to unseen systems. The proposed architecture enhances the efficiency of Boltzmann Generators trained on single molecular systems. The paper discusses the challenges of generating samples from equilibrium Boltzmann distributions in statistical physics, which traditionally require sequential sampling algorithms like Markov Chain Monte Carlo and Molecular Dynamics (MD) simulations. These methods are computationally expensive and time-consuming. Recent advances in machine learning have led to the development of Boltzmann Generators (BGs), which use normalizing flows to learn a transformation from a simple prior distribution to the target Boltzmann distribution. However, these BGs are typically limited to the system of interest and require retraining for new systems. The proposed TBGs overcome this limitation by being transferable across chemical space. They are trained on one set of molecules and can generalize to another set, allowing for efficient sampling without retraining. The framework is evaluated on dipeptides, where it demonstrates the ability to generate unbiased samples from unseen systems. The TBG architecture is shown to be effective in capturing metastable states and generating accurate free energy landscapes. The paper also discusses related work, including the use of normalizing flows and flow matching for training CNFs. It highlights the importance of equivariant flows in capturing the symmetries of molecular systems. The proposed TBG architecture is based on an equivariant graph neural network (EGNN) and is shown to be efficient in sampling and reweighting. The framework is evaluated on various dipeptides, including alanine, KS, and GN dipeptides, demonstrating its effectiveness in generating unbiased samples and accurate free energy projections. The paper concludes that transferable Boltzmann Generators can be effectively trained using small datasets and are capable of scaling to larger systems. The framework has the potential to accelerate drug and material discovery by replacing or enhancing MD simulations. However, the paper also acknowledges potential risks, including the possibility of misuse for identifying new diseases or developing biological weapons. The study highlights the importance of chemically informed priors and the role of network architecture in achieving accurate results.This paper introduces a framework for transferable Boltzmann Generators (TBGs) that can generate samples from unseen Boltzmann distributions without retraining. TBGs are based on continuous normalizing flows (CNFs) and are designed to be transferable across chemical space. The framework allows for efficient sampling from the target distribution and reweighting to the unbiased target Boltzmann distribution. The transferability of the framework is evaluated on dipeptides, where it shows efficient generalization to unseen systems. The proposed architecture enhances the efficiency of Boltzmann Generators trained on single molecular systems. The paper discusses the challenges of generating samples from equilibrium Boltzmann distributions in statistical physics, which traditionally require sequential sampling algorithms like Markov Chain Monte Carlo and Molecular Dynamics (MD) simulations. These methods are computationally expensive and time-consuming. Recent advances in machine learning have led to the development of Boltzmann Generators (BGs), which use normalizing flows to learn a transformation from a simple prior distribution to the target Boltzmann distribution. However, these BGs are typically limited to the system of interest and require retraining for new systems. The proposed TBGs overcome this limitation by being transferable across chemical space. They are trained on one set of molecules and can generalize to another set, allowing for efficient sampling without retraining. The framework is evaluated on dipeptides, where it demonstrates the ability to generate unbiased samples from unseen systems. The TBG architecture is shown to be effective in capturing metastable states and generating accurate free energy landscapes. The paper also discusses related work, including the use of normalizing flows and flow matching for training CNFs. It highlights the importance of equivariant flows in capturing the symmetries of molecular systems. The proposed TBG architecture is based on an equivariant graph neural network (EGNN) and is shown to be efficient in sampling and reweighting. The framework is evaluated on various dipeptides, including alanine, KS, and GN dipeptides, demonstrating its effectiveness in generating unbiased samples and accurate free energy projections. The paper concludes that transferable Boltzmann Generators can be effectively trained using small datasets and are capable of scaling to larger systems. The framework has the potential to accelerate drug and material discovery by replacing or enhancing MD simulations. However, the paper also acknowledges potential risks, including the possibility of misuse for identifying new diseases or developing biological weapons. The study highlights the importance of chemically informed priors and the role of network architecture in achieving accurate results.