KaLM-embedding-multilingual-mini-instruct-v2.5

<!-- <p align="center"> <img src="imgs/logo.jpg" width="800"/> <p> --> <h1 align="center">KaLM-Embedding-V2.5</h1> <p align="center"> <a href="https://huggingface.co/KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5"> <img src="https://img.shields.io/badge/%F0%9F%A4%97_HuggingFace-Model-ffbd45.svg" alt="HuggingFace"> </a> <a href="https://huggingface.co/spaces/nazib61/qdarnt"> <img src="https://img.shields.io/badge/%F0%9F%A4%97_HuggingFace-Demo-90EE90?style=flat" alt="Demo"> </a> <a href="https://github.com/HITsz-TMG/KaLM-Embedding"> <img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="GitHub Code"> </a> <a href="https://arxiv.org/abs/2506.20923"> <img src="https://img.shields.io/badge/Paper-KaLM--Embedding-d4333f?logo=arxiv&logoColor=white&colorA=cccccc&colorB=d4333f&style=flat" alt="Paper"> </a> </p>

Short Description

KaLM-Embedding-V2.5 is a versatile and compact embedding model, which achieves SOTA performance among models of comparable size and competes with models 3–26x larger by leveraging superior training techniques and data.

Model Details

• Model Size: 0.5B
• Embedding Dimension: 896
• Max Input Tokens: 32k
• MRL dimensions: 896, 512, 256, 128, and 64
• Attn: Bidirectional attention
• Pooling: Mean pooling

Training Recipe

• Large-scale weakly supervised pretraining
• High-quality supervised finetuning
• Contrastive distillation with fine-grained soft labels

Additionally, focal-style sample reweighting and online hard-negative mixing are employed to emphasize difficult samples and enrich hard negatives.

📑 Open-source Plan

• Model Checkpoint
  • KaLM-embedding-multilingual-mini-v1
  • KaLM-embedding-multilingual-mini-instruct-v1
  • KaLM-embedding-multilingual-mini-instruct-v1.5
  • KaLM-embedding-multilingual-mini-instruct-v2
  • KaLM-embedding-multilingual-mini-instruct-v2.5
  • KaLM-Embedding-Gemma3-12B-2511
• Training and Evaluation Code: HITsz-TMG/KaLM-Embedding
• Technical Report: KaLM-Embedding-V2: Superior Training Techniques and Data Inspire A Versatile Embedding Model
• Pre-training Data: Pre-training Data
• Fine-tuning Data: Fine-tuning Data

Evaluation

Overall results on MTEB (cmn, v1) and MTEB (eng, v1).

Detailed model performance on MTEB (cmn, v1).

Detailed model performance on MTEB (eng, v1).

OOD evaluation: KaLM-Embedding-V2.5 exhibits strong OOD generalization, competing with the 15x larger model in real-world retrieval scenarios.

Matryoshka embedding evaluation: KaLM-Embedding-V2.5 maintains robust performance with matryoshka embeddings even at smaller dimensions.

Requirements

Since we have used the Qwen2 model, we advise you to install transformers>=4.37.0, or you might encounter the following error:

KeyError: 'qwen2'

Usage

sentence-transformers support

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer(
    "KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5",
    trust_remote_code=True,
    model_kwargs={
        "torch_dtype": torch.bfloat16,
        "attn_implementation": "flash_attention_2",  # Optional
    },
)
model.max_seq_length = 512

sentences = ["This is an example sentence", "Each sentence is converted"]
embeddings = model.encode(
    sentences,
    normalize_embeddings=True,
    batch_size=256,
    show_progress_bar=True,
)
print(embeddings)
'''
[[-0.01043701 -0.02172852  0.0100708  ... -0.02807617  0.00157166
  -0.03637695]
 [-0.00424194  0.02966309  0.03686523 ... -0.02587891  0.01953125
  -0.00125122]]
'''

We add task instructions for asymmetric tasks: retrieval, reranking, classification, and clustering. And, we add task instructions for both queries and passages in symmetric tasks, including STS and pair classification. If you want to add task instructions to the query, you can use the model like this:

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer(
    "KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5",
    trust_remote_code=True,
    model_kwargs={
        "torch_dtype": torch.bfloat16,
        "attn_implementation": "flash_attention_2",  # Optional
    },
)
model.max_seq_length = 512

sentences = ["This is an example sentence", "Each sentence is converted"]
prompt = "Instruct: Classifying the category of french news.\nQuery:"
embeddings = model.encode(
    sentences,
    prompt=prompt,
    normalize_embeddings=True,
    batch_size=256,
    show_progress_bar=True,
)
print(embeddings)
'''
[[-0.01867676  0.02319336  0.00280762 ... -0.02075195  0.00196838
  -0.0703125 ]
 [-0.0067749   0.03491211  0.01434326 ... -0.0043335   0.00509644
  -0.04174805]]
'''

Or you can use encode_query and encode_document to automatically add the default prompt for queries ("Instruct: Given a query, retrieve documents that answer the query \n Query: ") and documents (""), respectively.

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer(
    "KaLM-Embedding/KaLM-embedding-multilingual-mini-instruct-v2.5",
    trust_remote_code=True,
    model_kwargs={
        "torch_dtype": torch.bfloat16,
        "attn_implementation": "flash_attention_2",  # Optional
    },
)
model.max_seq_length = 512

queries = [
    "What is the capital of China?",
    "Explain gravity",
]
documents = [
    "The capital of China is Beijing.",
    "Gravity is a force that attracts two bodies towards each other. It gives weight to physical objects and is responsible for the movement of planets around the sun.",
]

query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)

similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
'''
tensor([[0.9034, 0.2563],
        [0.3153, 0.7396]])
'''

vllm support

pip install -U vllm==0.8.5

import torch
import vllm
from vllm import LLM
def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery:{query}'

task = 'Given a query, retrieve documents that answer the query'
queries = [
    get_detailed_instruct(task, 'What is the capital of China?'),
    get_detailed_instruct(task, 'Explain gravity')
]
documents = [
    "The capital of China is Beijing.",
    "Gravity is a force that attracts two bodies towards each other. It gives weight to physical objects and is responsible for the movement of planets around the sun."
]
input_texts = queries + documents

model = LLM(model="{MODEL_NAME_OR_PATH}", task="embed", trust_remote_code=True, dtype="float16")

outputs = model.embed(input_texts)
embeddings = torch.tensor([o.outputs.embedding for o in outputs])
scores = (embeddings[:2] @ embeddings[2:].T)
print(scores.tolist())

Citation

If you find this model useful, please consider giving a star and citation.

@inproceedings{
      zhao2026kalmembeddingv2,
      title={Ka{LM}-Embedding-V2: Superior Training Techniques and Data Inspire A Versatile Embedding Model},
      author={Xinping Zhao and Xinshuo Hu and Zifei Shan and Shouzheng Huang and Yao Zhou and Xin Zhang and Zetian Sun and zhenyu liu and Dongfang Li and Xinyuan Wei and Youcheng Pan and Yang Xiang and Meishan Zhang and Haofen Wang and Jun Yu and Baotian Hu and Min Zhang},
      booktitle={The Fourteenth International Conference on Learning Representations},
      year={2026},
      url={https://openreview.net/forum?id=Y7qzhvWhcz}
}

@misc{hu2025kalmembedding,
      title={KaLM-Embedding: Superior Training Data Brings A Stronger Embedding Model}, 
      author={Xinshuo Hu and Zifei Shan and Xinping Zhao and Zetian Sun and Zhenyu Liu and Dongfang Li and Shaolin Ye and Xinyuan Wei and Qian Chen and Baotian Hu and Haofen Wang and Jun Yu and Min Zhang},
      year={2025},
      eprint={2501.01028},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.01028}, 
}

Contact

If you encounter any issue, feel free to contact us via the email: [email protected], [email protected]