Tom Aarsen
Extend the README snippet to use encode_query/encode_document, default to document
9508e3c | license: cc-by-nc-4.0 | |
| language: | |
| - en | |
| - multilingual | |
| base_model: | |
| - Qwen/Qwen3-4B | |
| pipeline_tag: feature-extraction | |
| tags: | |
| - finance | |
| - legal | |
| - healthcare | |
| - code | |
| - stem | |
| - medical | |
| - multilingual | |
| library_name: sentence-transformers | |
| <img src="https://i.imgur.com/oxvhvQu.png"/> | |
| # Releasing zeroentropy/zembed-1 | |
| In retrieval systems, [embedding models determine the quality of your search](https://www.zeroentropy.dev/articles/how-to-overcome-poor-search-results-with-the-right-embedding-solution). | |
| However, SOTA embedding models are closed-source and proprietary. At ZeroEntropy, we've trained a SOTA 4B open-weight multilingual embedding model that outperforms every competitor we benchmarked, and we're launching it here on HuggingFace. | |
| This model [outperforms](https://huggingface.co/zeroentropy/zembed-1#evaluations) `OpenAI text-embedding-large`, `Cohere Embed v4`, `gemini-embedding-001`, and `voyage-4-nano` across finance, healthcare, legal, conversational, manufacturing, code, and STEM. | |
| zembed-1 is distilled directly from our SOTA reranker [zerank-2](https://huggingface.co/zeroentropy/zerank-2) using our [zELO methodology](https://arxiv.org/abs/2509.12541), which models relevance scores as adjusted [Elo ratings](https://en.wikipedia.org/wiki/Elo_rating_system). Standard contrastive training on binary labels can't match this signal. See [our blog post](https://www.zeroentropy.dev/articles/introducing-zembed-1-the-worlds-best-multilingual-text-embedding-model) for details. | |
| The model supports flexible dimension projections (2560, 1280, 640, 320, 160, 80, 40) and quantization down to binary, compressing a full 8 KB vector to under 128 bytes with a controlled accuracy trade-off. See our Technical Report (Coming soon!) for details on the projection method. zembed-1 is multilingual from the ground up, with over half the training data in non-English languages. | |
| Since we're a small company, this model is only released under a non-commercial license. If you'd like a commercial license, please contact us at founders@zeroentropy.dev and we'll get you a license ASAP. | |
| ## How to Use | |
| ```python | |
| from sentence_transformers import SentenceTransformer | |
| # Initialize model | |
| model = SentenceTransformer( | |
| "zeroentropy/zembed-1", | |
| trust_remote_code=True, | |
| # truncate_dim=640, # Optional: Reduce dimensionality from 2560 to {1280, 640, 320, 160, 80, 40} | |
| model_kwargs={"torch_dtype": "bfloat16"}, | |
| ) | |
| # Define query and documents | |
| query = "What is backpropagation?" | |
| documents = [ | |
| "Backpropagation is a fundamental algorithm for training neural networks by computing gradients.", | |
| "Gradient descent is used to optimize model parameters during the training process.", | |
| "Neural network training relies on efficient computation of derivatives through backpropagation.", | |
| ] | |
| # Encode query and documents (uses task-specific prompts automatically) | |
| query_embeddings = model.encode_query(query) | |
| document_embeddings = model.encode_document(documents) | |
| # (2560,) (3, 2560) | |
| # Compute cosine similarities | |
| similarities = model.similarity(query_embeddings, document_embeddings) | |
| # tensor([[0.7525, 0.5670, 0.6835]]) | |
| ``` | |
| The model can also be used through ZeroEntropy's [/models/embed](https://docs.zeroentropy.dev/api-reference/models/embed) endpoint. | |
| ## Evaluations | |
| NDCG@10 scores between `zembed-1` and competing embedding models, averaged across public and private benchmarks per domain. Full per-benchmark breakdown [here](https://docs.google.com/spreadsheets/d/1qFXGZLMg6-O5tVLIJS3tpf5QNJxCHiiQtj35dZub4vY/edit?gid=0#gid=0). | |
| | Domain | ZeroEntropy zembed-1 | voyage-4-nano | Qwen3 4B | Cohere Embed v4 | gemini-embed-001 | jina-v5-small | OpenAI Large | bge-m3 | | |
| |------------------|----------------------|---------------|----------|-----------------|-------------------|---------------|--------------|--------| | |
| | Finance | **0.4476** | 0.4227 | 0.3715 | 0.3670 | 0.3291 | 0.3576 | 0.3291 | 0.3085 | | |
| | Healthcare | **0.6260** | 0.5356 | 0.5134 | 0.4750 | 0.5008 | 0.5132 | 0.5315 | 0.3620 | | |
| | Legal | **0.6723** | 0.5957 | 0.5858 | 0.5894 | 0.6069 | 0.5716 | 0.5099 | 0.5207 | | |
| | Conversational | **0.5385** | 0.4045 | 0.4034 | 0.4244 | 0.4247 | 0.4430 | 0.3988 | 0.3296 | | |
| | Manufacturing | **0.5556** | 0.4857 | 0.4932 | 0.4919 | 0.4664 | 0.4725 | 0.4736 | 0.3736 | | |
| | Web Search | 0.6165 | 0.5977 | 0.6914 | **0.7242** | 0.5881 | 0.6772 | 0.6750 | 0.6311 | | |
| | Code | **0.6452** | 0.6415 | 0.6379 | 0.6277 | 0.6305 | 0.6354 | 0.6155 | 0.5584 | | |
| | STEM & Math | **0.5283** | 0.5012 | 0.5219 | 0.4698 | 0.4840 | 0.3780 | 0.3905 | 0.3399 | | |
| | Enterprise | **0.3750** | 0.3600 | 0.2935 | 0.2915 | 0.3224 | 0.3012 | 0.3307 | 0.2213 | | |
| | **Average** | **0.5561** | **0.5050** | **0.5013** | **0.4957** | **0.4837** | **0.4833** | **0.4727** | **0.4050** | | |
| <img src="assets/zembed_eval_chart.png" alt="Bar chart comparing zembed-1 NDCG@10 scores against competing embedding models across domains" width="1000"/> |