KaLM-Embedding-V2.5

SOTA Multilingual Embedding Model from KaLM-Embedding

KaLM-Embedding-V2.5 is a lightweight multilingual embedding model with 0.5B parameters, achieving state-of-the-art performance among models of comparable size. It supports semantic search, retrieval-augmented generation, text similarity, clustering, and classification.

Key Features

• Lightweight: A compact 0.5B-parameter model.
• Multilingual: Multilingual text embeddings with strong English and Chinese performance.
• Adjustable dimensions: Choose 896, 512, 256, 128, or 64 dimensions through Matryoshka representation learning.
• Instruction-aware: Customize query embeddings for different retrieval tasks.
• Long inputs: Supports up to 32,768 tokens per formatted text. This API defaults to 512 tokens.
• Open source: Released under the Apache 2.0 license.

Usage

Provide an array of strings through texts. The API returns one embedding per text, preserving the input order.

For retrieval:

1. Encode queries with a task instruction.
2. Encode documents with an empty instruction.
3. Compare query and document embeddings using cosine similarity.

Use the same embedding dimension when encoding queries and documents.

Encode a Query

{
  "texts": [
    "What is the capital of France?"
  ],
  "instruction": "Given a question, retrieve passages that answer the question.",
  "embedding_dim": 896,
  "batch_size": 32,
  "max_length": 512,
  "normalize": true
}

Enter only the task description in instruction. The API automatically applies the instruction template.

Encode Documents

{
  "texts": [
    "Berlin is the capital of Germany.",
    "Paris is the capital of France.",
    "Rome is the capital of Italy.",
    "Madrid is the capital of Spain."
  ],
  "instruction": "",
  "embedding_dim": 896,
  "batch_size": 32,
  "max_length": 512,
  "normalize": true
}

Output

Each result contains:

• index: The zero-based position of the text in the input array.
• embedding: The embedding vector for that text.

With embedding_dim set to 896, each vector contains 896 values. With normalization enabled, dot products between query and document embeddings give cosine similarity scores.

Parameters

Smaller embedding dimensions reduce vector storage and similarity computation costs. Evaluate the quality trade-off on your own data.

Inputs exceeding max_length are truncated. Longer inputs and larger batches require more memory.

Model Details

References

• Hugging Face Model Card and Evaluation Results
• KaLM-Embedding-V2 Technical Report
• Replicate Serving Code
• KaLM-Embedding on Hugging Face
• Apache 2.0 License