spaCy Transformers is the official spaCy library from Explosion for plugging pretrained transformer models โ BERT, XLNet, GPT-2, and anything else on the Hugging Face hub โ into spaCy pipelines. The open-source project has around 1,400 stars on GitHub and remains maintained in 2026.
The library connects Hugging Face Transformers to spaCy’s pipeline architecture, so transformer embeddings can feed spaCy’s downstream components for named entity recognition, text classification, part-of-speech tagging, and other tasks. It handles the fiddly parts of the marriage: aligning subword tokenization between the transformer and spaCy’s tokenizer, batching documents efficiently, supporting shared transformer encoders across multiple pipeline components so the expensive model runs once, and enabling gradual fine-tuning of the transformer during training. Because it is built on spaCy’s configuration and training system, experiments are reproducible and models package into deployable pipelines the way other spaCy projects do. For many production NLP teams, this library is the standard route to transformer-quality accuracy while keeping spaCy’s speed, data structures, and tooling. It is released under the MIT license and works with the full range of Hugging Face model architectures, including multilingual encoders, so the same pipeline design carries across languages and domains without reworking the integration layer each time a better base model appears.
Free. spaCy Transformers is fully open source under the MIT license, with no paid tier.
NLP engineers using spaCy in production who want transformer-model accuracy for entity recognition, classification, and tagging within their existing pipelines.