Pre-trained RDF2vec models of all of Wikidata
To train the models, we used the GPU implementation of RDF2vec. Training was done on 8 H200 GPUs with 90GB/GPU. The largest model (200 dimensions) took 17.6 hours of training time.
The models are based on the Wikidata truthy dump from April 26th, 2026.
We provide three different flavors: embeddings trained with gensim, embeddings trained with the gpuRDF2vec PyTorch implementation, and embeddings trained with the grpuRDF2vec Pytorch Implementation after removing duplicate walks. Each of the flavors is available as a 100 and a 200 dimensional model. All models can be downloaded here:
For questions about the Wikidata embeddings, please get in touch with Martin Böckling.