README.md

---
license: other
language:
- zh
pipeline_tag: sentence-similarity
library_name: transformers
tags:
- transformers
- sentence-transformers
- sentence-similarity
- feature-extraction
- text-embeddings-inference
extra_gated_eu_disallowed: true
---
<p align="center">
    <img src="images/youtu_embedding.png" width="400"/>
<p>

<p align="center">
    🤗&nbsp;<a href="https://huggingface.co/tencent/Youtu-Embedding"><b>Hugging Face</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;
    🖥️&nbsp;<a href="https://github.com/TencentCloudADP/youtu-embedding"><b>GitHub</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;
    🌎&nbsp;<a href="https://arxiv.org/abs/2508.11442"><b>Technical Report</b></a>
</p>
<p align="center">
    💬&nbsp;<a href="https://huggingface.co/tencent/Youtu-Embedding/blob/main/images/wechat_qr.png"><b>WeChat</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;
    🤖&nbsp;<a href="https://discord.gg/QjqhkHQVVM"><b>Discord</b></a>
</p>

## 🎯 Introduction

**Youtu-Embedding** is a state-of-the-art, general-purpose text embedding model developed by Tencent Youtu Lab. It delivers exceptional performance across a wide range of natural language processing tasks, including Information Retrieval (IR), Semantic Textual Similarity (STS), Clustering, Reranking, and Classification.

  - **Top-Ranked Performance**: Achieved the #1 score of **77.58** on the authoritative CMTEB (Chinese Massive Text Embedding Benchmark) as of September 2025, demonstrating its powerful and robust text representation capabilities.

  - **Innovative Training Framework**: Features a Collaborative-Discriminative Fine-tuning Framework designed to resolve the "negative transfer" problem in multi-task learning. This is accomplished through a unified data format, task-differentiated loss functions, and a dynamic single-task sampling mechanism.


> **Note**: You can easily adapt and fine-tune the model on your own datasets for domain-specific tasks. For implementation details, please refer to the [training code](https://github.com/TencentCloudADP/youtu-embedding).


## 🤗 Model Download

| Model Name               | Parameters | Dimensions | Sequence Length | Download |
| :------------------- | :--------: | :--------: | :-----------------: | :------------------------------------------------------------------------------------------ |
| Youtu-Embedding   | 2B         | 2048       | 8K               | [Model](https://huggingface.co/tencent/Youtu-Embedding) |


## 🚀 Usage
#### 1. Using `transformers`
**📦 Installation**
```bash
pip install transformers==4.51.3
```
**⚙️ Usage**
```python
import torch
import numpy as np
from transformers import AutoModel, AutoTokenizer


class LLMEmbeddingModel():

    def __init__(self, 
                model_name_or_path, 
                batch_size=128, 
                max_length=1024, 
                gpu_id=0):
        self.model = AutoModel.from_pretrained(model_name_or_path, trust_remote_code=True)
        self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, padding_side="right")

        self.device = torch.device(f"cuda:{gpu_id}")
        self.model.to(self.device).eval()

        self.max_length = max_length
        self.batch_size = batch_size

        query_instruction = "Given a search query, retrieve passages that answer the question"
        if query_instruction:
            self.query_instruction = f"Instruction: {query_instruction} \nQuery:"
        else:
            self.query_instruction = "Query:"

        self.doc_instruction = ""
        print(f"query instruction: {[self.query_instruction]}\ndoc instruction: {[self.doc_instruction]}")

    def mean_pooling(self, hidden_state, attention_mask):
        s = torch.sum(hidden_state * attention_mask.unsqueeze(-1).float(), dim=1)
        d = attention_mask.sum(dim=1, keepdim=True).float()
        embedding = s / d
        return embedding
    
    @torch.no_grad()
    def encode(self, sentences_batch, instruction):
        inputs = self.tokenizer(
            sentences_batch,
            padding=True,
            truncation=True,
            return_tensors="pt",
            max_length=self.max_length,
            add_special_tokens=True,
        ).to(self.device)

        with torch.no_grad():
            outputs = self.model(**inputs)
            last_hidden_state = outputs[0]

            instruction_tokens = self.tokenizer(
                instruction,
                padding=False,
                truncation=True,
                max_length=self.max_length,
                add_special_tokens=True,
            )["input_ids"]
            if len(np.shape(np.array(instruction_tokens))) == 1:
                inputs["attention_mask"][:, :len(instruction_tokens)] = 0
            else:
                instruction_length = [len(item) for item in instruction_tokens]
                assert len(instruction) == len(sentences_batch)
                for idx in range(len(instruction_length)):
                    inputs["attention_mask"][idx, :instruction_length[idx]] = 0

            embeddings = self.mean_pooling(last_hidden_state, inputs["attention_mask"])
            embeddings = torch.nn.functional.normalize(embeddings, dim=-1)
        return embeddings

    def encode_queries(self, queries):
        queries = queries if isinstance(queries, list) else [queries]
        queries = [f"{self.query_instruction}{query}" for query in queries]
        return self.encode(queries, self.query_instruction)

    def encode_passages(self, passages):
        passages = passages if isinstance(passages, list) else [passages]
        passages = [f"{self.doc_instruction}{passage}" for passage in passages]
        return self.encode(passages, self.doc_instruction)

    def compute_similarity_for_vectors(self, q_reps, p_reps):
        if len(p_reps.size()) == 2:
            return torch.matmul(q_reps, p_reps.transpose(0, 1))
        return torch.matmul(q_reps, p_reps.transpose(-2, -1))

    def compute_similarity(self, queries, passages):
        q_reps = self.encode_queries(queries)
        p_reps = self.encode_passages(passages)
        scores = self.compute_similarity_for_vectors(q_reps, p_reps)
        scores = scores.detach().cpu().tolist()
        return scores


queries = ["What's the weather like?"]
passages = [
    'The weather is lovely today.',
    "It's so sunny outside!",
    'He drove to the stadium.'
]

model_name_or_path = "tencent/Youtu-Embedding"
model = LLMEmbeddingModel(model_name_or_path)
scores = model.compute_similarity(queries, passages)
print(f"scores: {scores}")
```

#### 2. Using `sentence-transformers`
**📦 Installation**
```bash
pip install sentence-transformers==5.1.0
```
**⚙️ Usage**
```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("tencent/Youtu-Embedding", trust_remote_code=True)
queries = ["What's the weather like?"]
passages = [
    'The weather is lovely today.',
    "It's so sunny outside!",
    'He drove to the stadium.'
]
queries_embeddings = model.encode_query(queries)
passages_embeddings = model.encode_document(passages)

similarities = model.similarity(queries_embeddings, passages_embeddings)
print(similarities)
```

#### 3. Using `LangChain` 🦜
Easily integrate the model into your **LangChain** applications, such as RAG pipelines.

**📦 Installation**

```bash
pip install langchain==0.3.27 langchain-community==0.3.29 langchain-huggingface==0.3.1 sentence-transformers==5.1.0 faiss-cpu==1.11.0
```

**⚙️ Usage**
```python
import torch
from langchain.docstore.document import Document
from langchain_community.vectorstores import FAISS
from langchain_huggingface.embeddings import HuggingFaceEmbeddings

model_name_or_path = "tencent/Youtu-Embedding"
device = "cuda" if torch.cuda.is_available() else "cpu"

model_kwargs = {
    'trust_remote_code': True,
    'device': device
}

embedder = HuggingFaceEmbeddings(
    model_name=model_name_or_path,
    model_kwargs=model_kwargs,
)

query_instruction = "Instruction: Given a search query, retrieve passages that answer the question \nQuery:"
doc_instruction = ""

data = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet."
]

documents = [Document(page_content=text, metadata={"id": i}) for i, text in enumerate(data)]
vector_store = FAISS.from_documents(documents, embedder, distance_strategy="MAX_INNER_PRODUCT")

query = "Which planet is known as the Red Planet?"
instructed_query = query_instruction + query
results = vector_store.similarity_search_with_score(instructed_query, k=3)

print(f"Original Query: {query}\n")
print("Results:")
for doc, score in results:
    print(f"- Text: {doc.page_content} (Score: {score:.4f})")

```

#### 4. Using `LlamaIndex` 🦙
This is perfect for integrating the model into your **LlamaIndex** search and retrieval systems.

**📦 Installation**

```bash
pip install llama-index==0.14.2 llama-index-embeddings-huggingface==0.6.1 sentence-transformers==5.1.0 llama-index-vector-stores-faiss==0.5.1
```

**⚙️ Usage**
```python
import faiss
import torch
from llama_index.core.schema import TextNode
from llama_index.core.vector_stores import VectorStoreQuery
from llama_index.vector_stores.faiss import FaissVectorStore
from llama_index.embeddings.huggingface import HuggingFaceEmbedding

model_name_or_path = "tencent/Youtu-Embedding"
device = "cuda" if torch.cuda.is_available() else "cpu"

embeddings = HuggingFaceEmbedding(
    model_name=model_name_or_path,
    trust_remote_code=True,
    device=device,
    query_instruction="Instruction: Given a search query, retrieve passages that answer the question \nQuery:",  
    text_instruction=""
)

data = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet."
]

nodes = [TextNode(id_=str(i), text=text) for i, text in enumerate(data)]

for node in nodes:
    node.embedding = embeddings.get_text_embedding(node.get_content())

embed_dim = len(nodes[0].embedding)
store = FaissVectorStore(faiss_index=faiss.IndexFlatIP(embed_dim))
store.add(nodes)

query = "Which planet is known as the Red Planet?"
query_embedding = embeddings.get_query_embedding(query)

results = store.query(
    VectorStoreQuery(query_embedding=query_embedding, similarity_top_k=3)
)

print(f"Query: {query}\n")
print("Results:")
for idx, score in zip(results.ids, results.similarities):
    print(f"- Text: {data[int(idx)]} (Score: {score:.4f})")

```


## 📊 CMTEB
| Model                    | Param.  | Mean(Task)         | Mean(Type)         | Class. | Clust. | Pair Class. | Rerank. | Retr.  | STS   |
| :------------------------ | :--------------| :----------------- | :----------------- | :----: | :----: | :---------: | :-----: | :----: | :---: |
| bge-multilingual-gemma2  | 9B | 67.64              | 68.52              | 75.31  | 59.30  | 79.30      | 68.28   | 73.73  | 55.19 |
| ritrieve\_zh\_v1       | 326M   | 72.71              | 73.85              | 76.88  | 66.50  | 85.98       | 72.86   | 76.97  | 63.92 |
| Qwen3-Embedding-4B      | 4B | 72.27              | 73.51              | 75.46  | 77.89  | 83.34       | 66.05   | 77.03  | 61.26 |
| Qwen3-Embedding-8B     | 8B | 73.84              | 75.00              | 76.97  | 80.08  | 84.23       | 66.99   | 78.21  | 63.53 |
| Conan-embedding-v2      | 1.4B | 74.24              | 75.99              | 76.47  | 68.84  | 92.44       | 74.41   | 78.31  | 65.48 |
| Seed1.6-embedding       | - | 75.63              | 76.68              | 77.98  | 73.11  | 88.71       | 71.65   | 79.69  | 68.94 |
| QZhou-Embedding         | 7B | 76.99              | 78.58              | 79.99  | 70.91  | 95.07       | 74.85   | 78.80  | 71.89 |
| **Youtu-Embedding** | 2B | **77.58** | **78.86** | 78.65 | 84.27 | 86.12 | 75.10 | 80.21 | 68.82 |

> **Note**: Comparative scores are from the MTEB [leaderboard](https://huggingface.co/spaces/mteb/leaderboard), recorded on September 28, 2025.


## 🎉 Citation
```bibtex
@misc{zhang2025codiemb,
  title={CoDiEmb: A Collaborative yet Distinct Framework for Unified Representation Learning in Information Retrieval and Semantic Textual Similarity},
  author={Zhang, Bowen and Song, Zixin and Chen, Chunquan and Zhang, Qian-Wen and Yin, Di and Sun, Xing},
  year={2025},
  eprint={2508.11442},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2508.11442},
}
```