BGE Embeddings (BAAI) - Chinese Embeddings

Basic Information

• Organization: BAAI (Beijing Academy of Artificial Intelligence)
• Country/Region: China (Beijing)
• Official Website: https://bge.baai.ac.cn
• GitHub: https://github.com/FlagOpen/FlagEmbedding
• Type: Open-source Text Embedding Model
• First Release: 2023
• Open Source License: MIT License

Product Description

BGE (BAAI General Embedding) is a series of general-purpose text embedding models developed by the Beijing Academy of Artificial Intelligence, serving as a benchmark product in the field of Chinese embeddings. The latest BGE-M3 model is renowned for its "multi-functionality, multi-language, and multi-granularity" features, capable of performing three common retrieval functions simultaneously: dense retrieval, multi-vector retrieval, and sparse retrieval. It supports over 100 working languages, making it a standout among open-source embedding models.

Core Features/Characteristics

BGE-M3 (Flagship Model)

• Three Retrieval Modes: Supports dense (Dense), multi-vector (Multi-vector/ColBERT), and sparse (Sparse) retrieval simultaneously
• 100+ Language Support: Extensive multilingual capabilities
• Long Document Support: Supports input up to 8192 tokens
• Parameter Scale: 568M parameters, compact yet powerful
• 1024-Dimensional Output: Embedding vector dimension is 1024

BGE-VL (Visual Language, Released March 2025)

• Multimodal Embedding: Supports visual search applications
• SOTA Performance: Leads in multimodal embedding benchmarks

BGE-en-ICL (In-Context Learning)

• In-Context Learning: Introduces in-context learning capabilities to embedding models
• Released July 2024: Innovative ICL embedding approach

BGE-multilingual-gemma2

• Based on Gemma-2-9B: Large-scale multilingual embedding model
• Multilingual SOTA: Achieves top performance in multilingual benchmarks

Model Matrix

Business Model

• Completely Free and Open Source: MIT License
• Free Deployment: Can be freely deployed and used in any environment
• No API Service: Primarily distributed as model weights
• Hugging Face: All models are publicly available on Hugging Face Hub

Target Users

• Chinese NLP application developers
• RAG system developers (especially for Chinese scenarios)
• Teams requiring local deployment of embedding models
• Multilingual retrieval system developers
• Academic researchers

Competitive Advantages

• Best-in-class Chinese embedding performance, far surpassing commercial models like OpenAI
• Completely open source and free (MIT license), with no usage restrictions
• BGE-M3's three retrieval modes are unique, offering flexible adaptation to different scenarios
• Supports long documents up to 8192 tokens
• Compact model (568M) with reasonable deployment resource requirements
• Active research team continuously releasing new models
• Excellent performance in MTEB Chinese and multilingual benchmarks

Limitations

• Requires self-deployment and management, with no hosted API
• Somewhat challenging for non-technical users
• Compared to ultra-large models (e.g., BGE-multilingual-gemma2 based on 9B parameters), the M3 model may be slightly weaker in certain tasks
• Lacks commercial-grade SLA and technical support

Relationship with the OpenClaw Ecosystem

BGE Embeddings is the preferred embedding model for OpenClaw's Chinese users. As a completely open-source local model, BGE perfectly aligns with OpenClaw's requirements for privacy protection and localization. BGE-M3's three retrieval modes (dense + sparse + multi-vector) provide OpenClaw with the most flexible retrieval strategies. For OpenClaw's Chinese knowledge base scenarios, BGE's Chinese comprehension capabilities far exceed those of English-first commercial models like OpenAI.