opensearch-neural-embed-v1
by opensearch-project
OpenSearch Project's first open embedding model: 598M parameters, 2048-dimension English text vectors, trained by distillation with no labelled data
opensearch-project/opensearch-neural-embed-v1Deploy opensearch-neural-embed-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
opensearch-neural-embed-v1 is a dense text embedding model for English from the OpenSearch Project, published on Hugging Face on 4 November 2026 under the Apache 2.0 license. It maps queries and documents to 2048-dimension, L2-normalized vectors compared by cosine similarity, for semantic search, passage retrieval and RAG. It starts from Qwen3-Embedding-0.6B and was distilled from two larger embedding models using about 17 million unlabeled English texts, with no relevance labels, no hard-negative mining and no contrastive fine-tuning. Queries take an instruction prompt; documents are encoded without one.
Architecture
A Sentence Transformers bi-encoder: the Qwen3-Embedding-0.6B backbone (596M parameters) with mean pooling and a 2.1M-parameter projection head to 2048 dimensions. Training ran in two stages with three objectives, pulling each student vector toward the teachers' vector for the same text, reproducing the teachers' in-batch similarity matrix, and keeping the teachers' ranking of text pairs. The maximum sequence length is 8,192 tokens; training used inputs up to 1,024 tokens.
Mixpeek SDK Integration
# Encode documents without a prompt (usage from the model card).
# pip install -U sentence-transformers
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("opensearch-project/opensearch-neural-embed-v1")
documents = [
"Transcript, 00:12:40: the presenter compares the two lenses side by side.",
"Slide 14: quarterly revenue by region, with EMEA up 12 percent.",
]
doc_vectors = model.encode_document(documents) # 2048-d, L2-normalized
# Store each vector with its file, timestamp and text in your vector index.Capabilities
- English semantic search and passage retrieval
- Query embeddings with a task instruction, document embeddings without one
- Inputs up to 8,192 tokens
- 2048-dimension vectors, half the size of Qwen3-Embedding-8B's 4096
- Runs in Sentence Transformers or plain Transformers with mean pooling
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MTEB (eng, v2) | Mean (Task) | 73.76 | OpenSearch Project model card; Qwen3-Embedding-0.6B 70.47 and Qwen3-Embedding-8B 75.23 on the same table |
| BEIR | Average NDCG@10 | 0.608 | OpenSearch Project model card |
Performance
We have not measured it. Both scores are from the OpenSearch Project's own evaluation; the comparison models' scores on the card come from the MTEB leaderboard or their own cards. Score it on your own queries before switching, and send queries through encode_query so the instruction prompt is applied.
Common Pipeline Companions
Frequently Asked Questions
What is opensearch-neural-embed-v1?
An English text embedding model from the OpenSearch Project, published on 4 November 2026 under Apache 2.0. It has 598M parameters, reads up to 8,192 tokens and returns 2048-dimension vectors for semantic search and retrieval. The card reports 73.76 on MTEB (eng, v2) and 0.608 average NDCG@10 on BEIR.
How was opensearch-neural-embed-v1 trained without labelled data?
By distillation. It starts from Qwen3-Embedding-0.6B and learns to reproduce two larger embedding models' vectors, similarity structure and rankings over about 17 million unlabeled English texts. The card lists no relevance labels, no hard-negative mining and no contrastive fine-tuning.
Do I need a prompt for queries?
Yes, for queries only. The card's default query prompt is "Instruct: Given a web search query, retrieve relevant passages that answer the query" followed by "Query:", which Sentence Transformers applies through encode_query. Encode documents with encode_document and no prompt.
Does opensearch-neural-embed-v1 only work with OpenSearch?
No. Its vectors are ordinary 2048-dimension float vectors compared by cosine similarity, so any vector store that indexes that size can hold them. OpenSearch is the project that trained it, and its k-NN indexes are one place to use it.
Can opensearch-neural-embed-v1 search images or video?
Not directly, since it embeds English text only. Index the text that comes out of your media, such as transcripts, captions and OCR, and use a multimodal embedding model to search the pixels themselves.
Specification
Build a pipeline with opensearch-neural-embed-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