This article introduces a transferable fermionic neural wavefunction for molecules, which maps uncorrelated Hartree-Fock orbitals to correlated neural network orbitals. The proposed ansatz enables learning a single wavefunction across multiple compounds and geometries, as demonstrated by transferring a pretrained model from smaller fragments to larger compounds. The model achieves high accuracy for relative energies and outperforms conventional methods like CCSD(T) with significantly less computational effort during fine-tuning. The approach is inspired by foundation models in language and vision, where extensive pre-training allows efficient transfer to new tasks. The model is shown to generalize well across chemically similar compounds and outperforms other methods like GLOBE in terms of accuracy and transferability. The ansatz is designed to be parameter-agnostic to system size, enabling efficient computation of potential energy surfaces (PES) across different compounds. The model is also equivariant to the sign of Hartree-Fock orbitals, ensuring robustness during pre-training. The study demonstrates that pre-training on a diverse dataset of 360 geometries leads to a foundation wavefunction model that can be fine-tuned to achieve high-accuracy ab-initio results with minimal computational effort. The model's performance is validated on various test systems, including hydrogen chains, methane, ethene, and cyclobutadiene, showing strong generalization and accuracy. The results highlight the potential of transferable neural wavefunctions for efficient and accurate quantum mechanical property predictions.This article introduces a transferable fermionic neural wavefunction for molecules, which maps uncorrelated Hartree-Fock orbitals to correlated neural network orbitals. The proposed ansatz enables learning a single wavefunction across multiple compounds and geometries, as demonstrated by transferring a pretrained model from smaller fragments to larger compounds. The model achieves high accuracy for relative energies and outperforms conventional methods like CCSD(T) with significantly less computational effort during fine-tuning. The approach is inspired by foundation models in language and vision, where extensive pre-training allows efficient transfer to new tasks. The model is shown to generalize well across chemically similar compounds and outperforms other methods like GLOBE in terms of accuracy and transferability. The ansatz is designed to be parameter-agnostic to system size, enabling efficient computation of potential energy surfaces (PES) across different compounds. The model is also equivariant to the sign of Hartree-Fock orbitals, ensuring robustness during pre-training. The study demonstrates that pre-training on a diverse dataset of 360 geometries leads to a foundation wavefunction model that can be fine-tuned to achieve high-accuracy ab-initio results with minimal computational effort. The model's performance is validated on various test systems, including hydrogen chains, methane, ethene, and cyclobutadiene, showing strong generalization and accuracy. The results highlight the potential of transferable neural wavefunctions for efficient and accurate quantum mechanical property predictions.