ModelsBAAIbge-m3-multi
providerBAAI /

bge-m3-multi

5 DZD in / 1M tokens

bge-m3-multi is an alternative designation for the bge-m3 embedding model developed by the Beijing Academy of Artificial Intelligence (BAAI). Built on a 567-million parameter XLM-RoBERTa architecture, the "Multi" suffix emphasizes its core design pillars: Multi-Linguality (supporting over 100 languages), Multi-Functionality (simultaneous Dense, Sparse, and Multi-Vector ColBERT outputs), and Multi-Granularity (handling inputs up to 512 tokens). It serves as a unified, state-of-the-art engine for hybrid search and complex multilingual Retrieval-Augmented Generation (RAG) pipelines.

Public
ProjectPaperLicense
bge-m3-multi
ArchitectureTransformer
Context Window512

1. Overview

  • Model Name: bge-m3
  • Developer: BAAI (Beijing Academy of Artificial Intelligence)
  • Model Type: Advanced Text Embedding Model
  • The "M3" Core Concept:
    • Multi-Lingual: Native support for over 100 languages.
    • Multi-Function: Supports Dense, Sparse (Lexical), and Multi-vector (ColBERT-style) retrieval.
    • Multi-Granularity: Handles input lengths from short sentences up to 512-token passages.

2. Technical Specifications

  • Architecture: Transformer-based (RetroMAE pre-training methodology)
  • Max Context Length: Up to 512 tokens
  • Embedding Dimension: 1024
  • Parameters: ~567 Million
  • Output Formats:
    1. Dense: Standard single vector representing the whole text.
    2. Sparse: Lexical weights (similar to BM25, good for keyword matching).
    3. Multi-Vector: Token-level embeddings (ColBERT-style for fine-grained alignment).

3. Performance & Capabilities

  • Cross-Lingual Retrieval: Highly optimized for matching queries in one language (e.g., Arabic) with documents in another (e.g., English).
  • Long-Document Processing: Accepts up to 512 tokens per input; longer inputs are truncated, so split long documents into chunks before embedding.

4. Hardware & Integration

  • Resource Requirements: Lightweight. Can run efficiently on CPUs for smaller batches, or entry-level GPUs (T4, L4) for high-throughput production.
  • Framework Compatibility: HuggingFace sentence-transformers, LlamaIndex, LangChain, FlagEmbedding.
  • Database Fit: Ideal for Vector DBs like PostgreSQL (Supabase pgvector), Pinecone, or Qdrant.

5. Primary Use Cases

  • High-accuracy RAG (Retrieval-Augmented Generation) pipelines.
  • Hybrid Search implementations (combining dense semantic search with sparse keyword search).
  • Multi-lingual enterprise search platforms.