mxbai-embed-large-v1
by mixedbread-ai
Mixedbread's 335M English text embedder: 1,024-dimension vectors, a fixed query prompt for retrieval, and support for shorter Matryoshka vectors and binary quantization
mixedbread-ai/mxbai-embed-large-v1Deploy mxbai-embed-large-v1
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
mxbai-embed-large-v1 is an English text embedding model from Mixedbread, published on Hugging Face in March 2024 under the Apache-2.0 license. It is a 24-layer BERT-large encoder with 335M parameters that turns a sentence or passage of up to 512 tokens into a 1,024-dimension vector. For retrieval, queries take the prompt "Represent this sentence for searching relevant passages: " and documents take no prompt.
The card reports a 64.68 average across 56 MTEB English datasets and 54.39 on its 15 retrieval datasets, against 64.23 and 54.29 for BAAI's bge-large-en-v1.5 and 64.58 and 55.44 for OpenAI's text-embedding-3-large. Mixedbread says the model was trained with no overlap with MTEB data. It is one of the most downloaded embedding models on Hugging Face, at about 2.1 million downloads a month.
Architecture
A BERT-large encoder (24 layers, hidden size 1,024, 512-token maximum input) with CLS pooling, which the card names as the default; mean pooling also works. Similarity is cosine. Trained with Matryoshka Representation Learning, so the first 512 or fewer dimensions can be kept as a shorter vector, and the embeddings hold up under int8 or binary quantization.
Mixpeek SDK Integration
# Embed documents with mxbai-embed-large-v1 (no prompt for documents).
# pip install -U sentence-transformers
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1", truncate_dim=512)
docs = ["A man is eating food.", "A man is riding a horse."]
doc_vecs = model.encode(docs) # 512 dims after Matryoshka truncation; drop truncate_dim for 1,024Capabilities
- 1,024-dimension English text embeddings
- Shorter vectors by truncation (Matryoshka), such as 512 dimensions
- int8 and binary quantization for smaller indexes
- Inputs up to 512 tokens
- Loads with Sentence Transformers, Transformers and Transformers.js
- Apache-2.0 license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MTEB English, 56 datasets | Average | 64.68 | Mixedbread model card, March 2024 |
| MTEB retrieval, 15 datasets | Average nDCG@10 | 54.39 | Mixedbread model card |
| MTEB STS, 10 datasets | Average | 85.00 | Mixedbread model card |
Performance
Longer passages are truncated at 512 tokens, so chunk documents before embedding. The card says retrieval queries must carry the query prompt. MTEB scores are from March 2024 and newer embedders score higher on the current leaderboard. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is mxbai-embed-large-v1?
An open English text embedding model from Mixedbread with 335M parameters, released under Apache-2.0 in March 2024. It produces 1,024-dimension vectors from inputs of up to 512 tokens and reports a 64.68 average on 56 MTEB English datasets.
What prompt does mxbai-embed-large-v1 need?
For retrieval, prefix each query with "Represent this sentence for searching relevant passages: " and embed documents with no prompt. In Sentence Transformers, encode(query, prompt_name="query") adds it.
Is mxbai-embed-large-v1 better than OpenAI text-embedding-3-large?
On the card's March 2024 MTEB table it averages 64.68 across 56 datasets against 64.58 for text-embedding-3-large, and 54.39 on retrieval against 55.44. The two are close; test both on a sample of your own queries before choosing.
Can I shrink mxbai-embed-large-v1 embeddings?
Yes. It was trained with Matryoshka Representation Learning, so you can keep the first 512 dimensions (truncate_dim=512), and it supports int8 and binary quantization. Binary vectors take 1 bit a dimension, 128 bytes for 1,024 dimensions.
How do I use mxbai-embed-large-v1 with Mixpeek?
Mixpeek's managed extractors do not include mxbai-embed-large-v1 today; running it as an extractor means registering a custom model, an Enterprise path. To compare it with your current search first, embed a sample of your documents as in the example on this page.
Specification
Build a pipeline with mxbai-embed-large-v1
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free