NeuronBlocks is a natural language processing deep learning toolkit from Microsoft that lets you build neural NLP models the way you would assemble Lego blocks — by configuration rather than by writing model code. The open-source project has around 1,500 stars on GitHub.
The toolkit decomposes NLP architectures into reusable blocks — embeddings, encoders, attention layers, decoders — that you combine through a JSON configuration file describing the model you want. With a model defined as configuration, NeuronBlocks handles training, evaluation, and prediction, and ships with implementations for the classic NLP task families: text classification, question answering, sequence labeling such as named entity recognition, text matching, and sentence-pair modeling. It also covers the production concerns around those models, including knowledge distillation and model compression to shrink trained networks for deployment. Built on PyTorch, it was designed for both researchers who want to prototype architectures quickly and engineers who need repeatable training pipelines. The repository’s most recent activity was in 2023, so today it serves best as a well-structured reference and a fast way to stand up standard NLP baselines rather than a cutting-edge LLM framework. It is released under the MIT license.
Free. NeuronBlocks is fully open source under the MIT license, with no paid tier or hosted offering.
NLP practitioners who want to assemble, train, and compress standard neural text models quickly from configuration instead of hand-writing architectures.