Model family
E5 Models
E5 models are widely used for multilingual embeddings, semantic search, retrieval, low-overhead indexing, and RAG pipelines.
Best for
Embedding
Use this family hub to compare E5 variants for embedding workflows, then open the detail page for deeper deployment notes.
RAG
Use this family hub to compare E5 variants for rag workflows, then open the detail page for deeper deployment notes.
Multilingual
Use this family hub to compare E5 variants for multilingual workflows, then open the detail page for deeper deployment notes.
Search
Use this family hub to compare E5 variants for search workflows, then open the detail page for deeper deployment notes.
Source box
This family currently includes 8 records tied to an exact published checkpoint and 1 broader or unresolved record. Identity is recorded per model so a representative checkpoint is never treated as the whole family.
Identity checked: 2026-08-21
Artifact identity does not establish licence or context truth. Those checks remain separate.
Variants
E5 models grouped by workflow
Embedding and reranking
Multilingual E5 Large
Microsoft / intfloat · E5
Best for: Teams building multilingual retrieval, semantic search, and RAG pipelines.
Local: Runs locally for many embedding and semantic search prototypes on CPU or modest GPU hardware.
e5-mistral-7b-instruct
Microsoft / intfloat · E5
Best for: Teams comparing larger instruction-tuned embedding models for retrieval quality when smaller E5 checkpoints are not enough.
Local: Task instructions go on the query side only ('Instruct: {task} Query: {q}'); documents are embedded without instructions. Embeddings come from last-token pooling, not CLS.
multilingual-e5-large-v2
Microsoft / intfloat · E5
Best for: Teams building multilingual search or RAG where one English-only embedding model is not enough.
Local: Use the exact checkpoint and quantization that matches your hardware and latency target.
e5-large-v2
Microsoft / intfloat · E5
Best for: Teams that want a strong English retrieval baseline for search and RAG before moving to multilingual or larger embedding experiments.
Local: Prefix every search string with 'query: ' and every document with 'passage: '; for symmetric similarity tasks the card says to use 'query: ' on both sides.
e5-base-v2
Microsoft / intfloat · E5
Best for: Builders who want a balanced English embedding model for search and RAG on moderate local or hosted infrastructure.
Local: 'query: ' on search strings, 'passage: ' on documents — the prefixes are part of the model's training contract, not optional formatting.
e5-small-v2
Microsoft / intfloat · E5
Best for: Local or cost-sensitive embedding pipelines that need a practical English retrieval model without the overhead of larger E5 releases.
Local: Same contract as its siblings: 'query: ' and 'passage: ' prefixes on everything, mean pooling, cosine similarity.
e5-large
Microsoft / intfloat · E5
Best for: Legacy large embedding baseline
Local: Existing indexes are not portable across versions: e5-large and e5-large-v2 vectors live in different spaces, so upgrading means re-embedding the corpus.
e5-base
Microsoft / intfloat · E5
Best for: Legacy base embedding baseline
Local: The E5 prefix rule predates v2 and applies here too; embeddings are mean-pooled and compared by cosine similarity. At this size the whole index pipeline — embed, store, search — fits in a container with no GPU drivers installed.
e5-small
Microsoft / intfloat · E5
Best for: Legacy lightweight embedding baseline
Local: Requires the 'query: '/'passage: ' prefixes like all E5 models; forgetting them degrades results silently.
Compare
All E5 models in the directory
| Model | Type | Best for | Local runner notes | License | Detail |
|---|---|---|---|---|---|
| Multilingual E5 Large | Embedding | Teams building multilingual retrieval, semantic search, and RAG pipelines. | Runs locally for many embedding and semantic search prototypes on CPU or modest GPU hardware. | MIT | Open |
| e5-mistral-7b-instruct | Embedding | Teams comparing larger instruction-tuned embedding models for retrieval quality when smaller E5 checkpoints are not enough. | Task instructions go on the query side only ('Instruct: {task} Query: {q}'); documents are embedded without instructions. Embeddings come from last-token pooling, not CLS. | MIT | Open |
| multilingual-e5-large-v2 | Embedding | Teams building multilingual search or RAG where one English-only embedding model is not enough. | Use the exact checkpoint and quantization that matches your hardware and latency target. | MIT | Open |
| e5-large-v2 | Embedding | Teams that want a strong English retrieval baseline for search and RAG before moving to multilingual or larger embedding experiments. | Prefix every search string with 'query: ' and every document with 'passage: '; for symmetric similarity tasks the card says to use 'query: ' on both sides. | MIT | Open |
| e5-base-v2 | Embedding | Builders who want a balanced English embedding model for search and RAG on moderate local or hosted infrastructure. | 'query: ' on search strings, 'passage: ' on documents — the prefixes are part of the model's training contract, not optional formatting. | MIT | Open |
| e5-small-v2 | Embedding | Local or cost-sensitive embedding pipelines that need a practical English retrieval model without the overhead of larger E5 releases. | Same contract as its siblings: 'query: ' and 'passage: ' prefixes on everything, mean pooling, cosine similarity. | MIT | Open |
| e5-large | Embedding | Legacy large embedding baseline | Existing indexes are not portable across versions: e5-large and e5-large-v2 vectors live in different spaces, so upgrading means re-embedding the corpus. | MIT | Open |
| e5-base | Embedding | Legacy base embedding baseline | The E5 prefix rule predates v2 and applies here too; embeddings are mean-pooled and compared by cosine similarity. At this size the whole index pipeline — embed, store, search — fits in a container with no GPU drivers installed. | MIT | Open |
| e5-small | Embedding | Legacy lightweight embedding baseline | Requires the 'query: '/'passage: ' prefixes like all E5 models; forgetting them degrades results silently. | MIT | Open |