E5
MicrosoftE5 embeds queries and passages for retrieval, trained by contrastive pre-training on weakly supervised text pairs and then on labelled retrieval data. The multilingual checkpoints initialize from XLM-RoBERTa and cover 100 languages at a 512-token limit, while e5-mistral-7b-instruct is a decoder-based, instruction-conditioned variant with a longer context. Every input carries an explicit "query:" or "passage:" prefix. A Microsoft Research team publishes the weights under the intfloat account.
The declared repository is the microsoft/unilm monorepo, whose e5 subtree carries the project; it holds evaluation scripts rather than a training pipeline.
Openness
3 high confidence- weights
- open(MIT ungated safetensors and PyTorch checkpoints across the English, multilingual and Mistral-based tiers)
- data
- components-listed(both stages are itemized dataset by dataset with a link per entry and pair counts, nine weak-supervision sources and twelve supervised ones
- code
- partial(the unilm/e5 subtree publishes MTEB and BEIR evaluation scripts and the cards give inference code
- license
- MIT(OSI)
An MIT release with an unusually well-documented corpus and no way to rebuild it. The card names every weak-supervision source with its pair count, which is more than most of this category offers, but the filtering pipeline is described only in the paper and the public repository stops at evaluation.
- https://huggingface.co/api/models/intfloat/e5-mistral-7b-instruct recorded 2026-09-11
license: mit; gated: false; private: false; the 7B decoder checkpoint is distributed under the same MIT terms.
- https://huggingface.co/api/models/intfloat/multilingual-e5-large recorded 2026-09-11
license: mit; gated: false; private: false; downloadable checkpoint files present.
- https://huggingface.co/intfloat/multilingual-e5-large/raw/main/README.md recorded 2026-09-11
Itemizes the first-stage weak-supervision mixture - filtered mC4 1B pairs, CC News 400M, NLLB 2.4B, Wikipedia 150M, filtered Reddit 800M, S2ORC 100M, StackExchange 50M, xP3 80M - without releasing the filtered corpus or the filtering scripts.
- https://raw.githubusercontent.com/microsoft/unilm/master/e5/README.md recorded 2026-09-11
The e5 subtree documents eval_mteb_beir.sh and eval_mteb_except_retrieval.sh and a synthetic passkey dataset; it publishes no training entry point.
Adoption
5 high confidenceAdoption is measured as Hugging Face downloads across the 13 shipped E5 checkpoints in the intfloat account, led by the multilingual-e5-small and multilingual-e5-large encoders.
- https://huggingface.co/api/models/intfloat/e5-mistral-7b-instruct recorded 2026-09-11
downloads: 221159.
- https://huggingface.co/api/models/intfloat/multilingual-e5-large recorded 2026-09-11
downloads: 7002947.
- https://huggingface.co/api/models/intfloat/multilingual-e5-small recorded 2026-09-11
downloads: 12369093.
Capability
3 high confidenceE5 is a workhorse embedding model, not a frontier one: its MTEB average matches today's small encoders, but the multilingual tier stops at roughly 512 tokens and the family has shipped no newer generation, unlike the longer-context, more recent models now ahead of it.
- https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base/raw/main/README.md recorded 2026-09-11
MTEB leaderboard table lists multilingual-e5-large-instruct at 560M parameters, 1024 dimensions, 514 maximum sequence length and an MTEB English average of 64.41.
- https://huggingface.co/intfloat/multilingual-e5-large/raw/main/README.md recorded 2026-09-11
Documents 100-language support from the XLM-RoBERTa initialization and tokenizes usage examples at max_length=512.
Verified 2026-09-11