Multilingual E5 Large Instruct
intfloat/multilingual-e5-large-instruct
published Feb 2024 · updated Jul 2025
Multilingual E5 Large Instruct is a feature-extraction model that generates multilingual text embeddings for tasks like retrieval, using instruction-tuned query prefixes.
specs
| Task | Feature Extraction |
| Architecture | Transformer (xlm-roberta-large) |
| Embedding Dimension | 1024 |
| Max Tokens | 512 |
| Languages | 100 |
about this model
intfloat/multilingual-e5-large-instruct is a feature-extraction (text embedding) model that produces 1024-dimensional multilingual embeddings and is optimized for retrieval tasks via instruction tuning. It is initialized from xlm-roberta-large and supports 100 languages, with performance degradation possible for low-resource languages.
Training and Usage
The model was trained in two stages: contrastive pre-training on 1 billion weakly supervised multilingual text pairs, followed by fine-tuning on labeled datasets from the E5-mistral paper. For retrieval, each query must be prefixed with a one-sentence task instruction (e.g., "Given a web search query, retrieve relevant passages that answer the query"); documents do not require an instruction. Pooling is performed via average pooling of the last hidden states, followed by L2 normalization.
Benchmark Performance
On the BEIR benchmark, multilingual-e5-large-instruct achieves a score of 52.5, the highest among the multilingual E5 model family:
| Model | BEIR Score |
|---|---|
| multilingual-e5-small | 46.6 |
| multilingual-e5-base | 48.9 |
| multilingual-e5-large | 51.4 |
| multilingual-e5-large-instruct | 52.5 |
The instruction-tuned model is reported to be on par with state-of-the-art English-only models of similar size. Detailed evaluation instructions for each task are defined in the official E5 repository.
Limitations
Input texts longer than 512 tokens are truncated. The model uses a low temperature (0.01) for InfoNCE loss, resulting in cosine similarity scores typically between 0.7 and 1.0; only the relative order of scores is meaningful.
best for
- ·Multilingual semantic search and passage retrieval
- ·Cross-lingual text classification
- ·Text similarity and clustering across 100 languages
FAQ
Yes, prepend a one-sentence task description to each query. Documents do not need instructions.
Differences in versions of transformers and PyTorch can cause negligible non-zero performance differences.
The model uses a low temperature of 0.01 for InfoNCE contrastive loss; relative order matters, not absolute values.
Long texts are truncated to at most 512 tokens.
Use the gigarouter OpenAI-compatible endpoint with your API key.
# OpenAI client - just change base_url from openai import OpenAI client = OpenAI(base_url="https://gigarouter.ai/v1", api_key=KEY) v = client.embeddings.create(model="intfloat/multilingual-e5-large-instruct", input=["hello world"]) print(v.data[0].embedding[:4])
try it live
runs the real hosted model on a shared demo allowance · get your own key + $25 free →