multilingual-e5-small (F16)

intfloat/multilingual-e5-small stored in float16 instead of float32: half the download (235 MB instead of 471 MB), same outputs.

Every weight of the original checkpoint is an exact float16 value, so the conversion is lossless: loading this file as float32 reproduces the original tensors bit for bit.

tokenizer.json and config.json are unchanged from the original repository.

Usage is the same as the original model: prefix queries with query: and documents with passage: , mean-pool the last hidden state and L2-normalize.

License

MIT, as the original model. Credit: Wang et al., Multilingual E5 Text Embeddings: A Technical Report (2024), https://arxiv.org/abs/2402.05672