multilingual-e5-large-instruct

Overview. multilingual-e5-large-instruct is an open-weight text embedding model from Microsoft, released under the MIT licence. It has 560M parameters and 24 layers, is initialised from xlm-roberta-large and trained with contrastive pre-training and fine-tuning. It returns a 1024-dimensional vector per input and reads up to 512 tokens. The model is described in 'Multilingual E5 Text Embeddings: A Technical Report' (Wang et al., February 2024).

Key strengths. One model covers the roughly 100 languages of xlm-roberta, so queries and documents in different languages land in the same vector space. Queries carry a short task instruction, which lets one model serve retrieval, classification and clustering. At 560M parameters it is small and cheap to run for large document collections.

When to use it. Use it for semantic search and retrieval-augmented generation over multilingual or mixed-language content, for cross-lingual search, and for clustering or classifying short texts. Write each query as 'Instruct: {task description}\nQuery: {query}', for example 'Given a web search query, retrieve relevant passages that answer the query'. Embed documents without an instruction.

Limitations. Input longer than 512 tokens is truncated, so long documents must be split into chunks. Quality drops for low-resource languages. Cosine similarity scores cluster between about 0.7 and 1.0, so rank results by relative score rather than a fixed threshold. Queries without the instruction prefix give weaker results.