MMdnn is a comprehensive set of tools from Microsoft that helps you work across different deep learning frameworks โ converting trained models between them and visualizing model architectures. It is open source on GitHub under the MIT license and has earned about 5,800 stars, with its latest updates in 2025.
Deep learning frameworks historically trapped models in their own formats: a network trained in Caffe could not easily run in TensorFlow or ship to a phone via CoreML. MMdnn solves this with an intermediate representation: it parses models from source frameworks into a common graph format, then emits them for target frameworks. It supports conversion among Caffe, Keras, MXNet, TensorFlow, CNTK, PyTorch (via ONNX) and CoreML, covering both model architecture and trained weights. Alongside the converter, MMdnn includes a visualizer that renders network graphs so you can inspect and compare architectures โ the popular “mmdnn” model visualization that many practitioners know from the web demo. For years it was a standard answer to the everyday problem of framework lock-in, particularly when deploying research models to mobile and edge runtimes, and its converters were also used to assemble public collections of ready-to-run converted networks.
MMdnn is completely free and open source under the MIT license.
MMdnn is for deep learning practitioners who need to move a model from the framework it was trained in to the framework or runtime they actually deploy on โ or who want a clear visual map of an unfamiliar architecture.