Dataflow models are a promising platform for programming language implementation due to their ability to extract the latent parallelism present in a given program. In spite of this, no modern-day language has emerged which leverages this property of implicit parallelism. As a first step towards the creation of such a language, we introduce a virtual machine design which is based on the tagged-token dataflow model. Notably, this design offers the fundamental concepts of the underlying model as first-class entities, which makes it possible to support various high-level language features without the need for any additional runtime support.