What changed
Hugging Face has released bekko-static-embedding-v1, a multilingual static embedding model that achieves state-of-the-art retrieval results. It scores about 11 points higher than other multilingual static models on multilingual retrieval tasks and about 2 points higher than English-only static models on English retrieval tasks. The model uses a custom tokenizer and was trained contrastively on about 814 million text pairs across 40 languages.
What this means in practice
The model is designed for fast retrieval without requiring GPU resources, running up to 400x faster on CPU than models like multilingual-e5-small. It can be used with Sentence Transformers to encode text in multiple languages, including English, French, Spanish, and Chinese. The model's performance is close to that of Transformer-based models, with results on MMTEB retrieval tasks within about 1 point of multilingual-e5-small, a 12-layer Transformer model.
Limitations and availability
The model is available for use with Sentence Transformers and can be loaded directly. It uses a custom 200k-token tokenizer and has 1024 dimensions. Truncating to 512 dimensions results in a minor drop in performance, costing about 0.9 points of average retrieval score. The model is intended for retrieval tasks and may not be suitable for other types of NLP tasks. The supplied announcement does not specify the exact release date or availability beyond the model being released.
