Virtual brain twins are personalized, generative, and adaptive brain models that use individual brain data for scientific and clinical purposes. They aim to mechanistically explain and capture the most relevant data features rather than replicate the biological brain in detail. These models are informed by subject-specific data and are used to guide decision-making in diagnostics, prognosis, and therapy. The models are composed of two main components: generative brain dynamics and an observer. The first component describes the evolution of neural activity over time and space, while the second component maps recorded signals from current neural activity. The models can be used to simulate brain activity and predict outcomes based on various conditions. The standard model of a virtual brain twin integrates various concepts and methods of virtual brain modeling over the past 20 years, emphasizing neural mass and neural field large-scale modeling, nonlinear dynamics, and network science. The model is defined by the Jirsa–Haken equation, which includes local node dynamics, local and global network interactions, and dynamical noise. The control parameters of the model are derived from the posterior distribution of the observed data, the brain dynamic model, and the forward solution. Personalized whole-brain network modeling relies on subject-specific parameters extracted from an individual's brain imaging data. Three levels of personalization are used: building a whole-brain model on the subject-specific brain space, directly mapping connectivity and other parameters into the brain models, and inferring clinically relevant parameters through model inversion or data fitting. The models are used in various clinical applications, including epilepsy, Alzheimer's disease, multiple sclerosis, Parkinson's disease, and psychiatric disorders. The virtual brain twin can be used to estimate epileptogenic networks, predict the outcome of surgical interventions, and simulate virtual stimulation. The models can also be used to study the dynamics of status epilepticus and the propagation of brain activity. In Alzheimer's disease, the models can be used to study the accumulation of amyloid-β plaques and neurofibrillary tangles, and to predict the progression of the disease. In multiple sclerosis, the models can be used to study the effects of structural lesions on conduction velocities and to predict clinical disability. In Parkinson's disease, the models can be used to study the effects of dopamine loss on neural activity and to predict the optimal stimulation paradigm. In psychiatric disorders, the models can be used to study the effects of neurotransmitter pathways on brain activity and to predict the optimal treatment. The key challenges in the development of virtual brain twins include degeneracy, overfitting, and the need for high-resolution models. Degeneracy refers to the ability of structurally different elements to produce the same function or behavior, which is a natural property of the brain. Overfitting occurs when a model is fitted to a specific dataset and performs exceptionally well on that dataset but fails to generalize to new data. High-resolution models are needed to improve the simulation and predictive power of the models. The future directions for virtual brain twinsVirtual brain twins are personalized, generative, and adaptive brain models that use individual brain data for scientific and clinical purposes. They aim to mechanistically explain and capture the most relevant data features rather than replicate the biological brain in detail. These models are informed by subject-specific data and are used to guide decision-making in diagnostics, prognosis, and therapy. The models are composed of two main components: generative brain dynamics and an observer. The first component describes the evolution of neural activity over time and space, while the second component maps recorded signals from current neural activity. The models can be used to simulate brain activity and predict outcomes based on various conditions. The standard model of a virtual brain twin integrates various concepts and methods of virtual brain modeling over the past 20 years, emphasizing neural mass and neural field large-scale modeling, nonlinear dynamics, and network science. The model is defined by the Jirsa–Haken equation, which includes local node dynamics, local and global network interactions, and dynamical noise. The control parameters of the model are derived from the posterior distribution of the observed data, the brain dynamic model, and the forward solution. Personalized whole-brain network modeling relies on subject-specific parameters extracted from an individual's brain imaging data. Three levels of personalization are used: building a whole-brain model on the subject-specific brain space, directly mapping connectivity and other parameters into the brain models, and inferring clinically relevant parameters through model inversion or data fitting. The models are used in various clinical applications, including epilepsy, Alzheimer's disease, multiple sclerosis, Parkinson's disease, and psychiatric disorders. The virtual brain twin can be used to estimate epileptogenic networks, predict the outcome of surgical interventions, and simulate virtual stimulation. The models can also be used to study the dynamics of status epilepticus and the propagation of brain activity. In Alzheimer's disease, the models can be used to study the accumulation of amyloid-β plaques and neurofibrillary tangles, and to predict the progression of the disease. In multiple sclerosis, the models can be used to study the effects of structural lesions on conduction velocities and to predict clinical disability. In Parkinson's disease, the models can be used to study the effects of dopamine loss on neural activity and to predict the optimal stimulation paradigm. In psychiatric disorders, the models can be used to study the effects of neurotransmitter pathways on brain activity and to predict the optimal treatment. The key challenges in the development of virtual brain twins include degeneracy, overfitting, and the need for high-resolution models. Degeneracy refers to the ability of structurally different elements to produce the same function or behavior, which is a natural property of the brain. Overfitting occurs when a model is fitted to a specific dataset and performs exceptionally well on that dataset but fails to generalize to new data. High-resolution models are needed to improve the simulation and predictive power of the models. The future directions for virtual brain twins