The Simformer is a new method for simulation-based amortized Bayesian inference that overcomes the limitations of existing approaches. It uses a combination of transformers and probabilistic diffusion models to estimate the joint distribution of parameters and data, enabling efficient inference for a wide range of tasks. Unlike previous methods that require fixed prior and simulator specifications, the Simformer can handle function-valued parameters, missing or unstructured data, and arbitrary conditionals of the joint distribution. It can also perform inference for observation intervals, making it highly flexible and efficient. The Simformer outperforms existing methods on benchmark tasks and is particularly effective in scenarios with complex dependencies. It can be applied to simulators in ecology, epidemiology, and neuroscience, and has demonstrated its ability to handle infinite-dimensional parameters and unstructured data. The Simformer provides an 'all-in-one' inference method that can sample all conditionals of the joint distribution, including posterior and likelihood. It also allows for the incorporation of domain knowledge through attention masks, improving simulation efficiency. The Simformer is accurate, flexible, and efficient, making it a promising tool for simulation-based inference in science and engineering.The Simformer is a new method for simulation-based amortized Bayesian inference that overcomes the limitations of existing approaches. It uses a combination of transformers and probabilistic diffusion models to estimate the joint distribution of parameters and data, enabling efficient inference for a wide range of tasks. Unlike previous methods that require fixed prior and simulator specifications, the Simformer can handle function-valued parameters, missing or unstructured data, and arbitrary conditionals of the joint distribution. It can also perform inference for observation intervals, making it highly flexible and efficient. The Simformer outperforms existing methods on benchmark tasks and is particularly effective in scenarios with complex dependencies. It can be applied to simulators in ecology, epidemiology, and neuroscience, and has demonstrated its ability to handle infinite-dimensional parameters and unstructured data. The Simformer provides an 'all-in-one' inference method that can sample all conditionals of the joint distribution, including posterior and likelihood. It also allows for the incorporation of domain knowledge through attention masks, improving simulation efficiency. The Simformer is accurate, flexible, and efficient, making it a promising tool for simulation-based inference in science and engineering.