BGE-M3 BAAI (2024年)（ビージーイーエムスリー）
2024年1月BAAI発表BGE-M3・Industry-leading multi-functionality dense+sparse+multi-vector + Industry-leading 100+ languages multilingual + Industry-leading 8192 max tokens + Industry-leading BAAI BGE-M3 MIT open-source multi-functionality embedding 2024。
概要
BGE-M3 BAAI は、2024年1月BAAI発表BGE-M3・Industry-leading multi-functionality dense+sparse+multi-vector 2024年BAAI BGE-M3 MIT open-source multi-functionality embedding position確立。BGE-M3 specifications = Industry-leading multi-functionality embedding (Industry-leading BGE-M3 BAAI 2024 + Industry-leading multi-functionality embedding + Industry-leading BAAI flagship) + Industry-leading 100+ languages multilingual + Industry-leading 8192 max tokens + Industry-leading BAAI BGE-M3 MIT open-source multi-functionality embedding 2024。
主な特徴・仕組み
- Developer: Industry-leading BAAI (Beijing Academy of Artificial Intelligence, Founded 2018 Beijing China)
- Model: Industry-leading BAAI BGE-M3 open-source multi-functionality embedding model 2024年1月 release
- Multi-Functionality: Industry-leading multi-functionality dense + sparse + multi-vector signature (M3 = Multi-Granularity, Multi-Lingual, Multi-Functionality)
- 100+ Languages: Industry-leading 100+ languages multilingual signature
- 8192 Max Tokens: Industry-leading 8192 max tokens signature (vs 512 typical)
- 1024d Dense: Industry-leading 1024d dense vector
- Sparse Vector: Industry-leading sparse vector lexical matching signature
- Multi-Vector ColBERT-Like: Industry-leading multi-vector ColBERT-like signature
- 66.4 BEIR Score: Industry-leading 66.4 BEIR retrieval score
- MTEB Top-Tier: Industry-leading MTEB top-tier signature
- HuggingFace + FlagEmbedding: Industry-leading HuggingFace + FlagEmbedding native
- MIT License: Industry-leading MIT license signature
- BAAI 2018 Beijing: Industry-leading BAAI 2018 Beijing China 6 years
- Industry-Leading: Industry-leading BGE-M3 BAAI MIT open-source multi-functionality embedding 2024
スペック比較表
| Embedding Model 2024 | Developer | Dimension | Pricing | Industry Position |
|---|---|---|---|---|
| BGE-M3 | BAAI | 1024 | Free open-source | Industry-leading BAAI BGE-M3 multi-functionality |
| OpenAI text-embedding-3-small | OpenAI | 256-1536 (Matryoshka) | $0.02/1M tokens | Industry-leading OpenAI efficient |
| Cohere embed-v3 | Cohere | 1024 (default) | $0.10/1M tokens | Industry-leading Cohere multilingual quality |
| Voyage AI voyage-3 | Voyage AI | 1024 (default) | $0.06/1M tokens | Industry-leading Voyage AI voyage-3 enterprise |
| Nomic Embed Text v1.5 | Nomic AI | 64-768 (Matryoshka) | Free open-source | Industry-leading Nomic Embed Text v1.5 Apache 2.0 |
具体例・対応製品
- BGE-M3 (2024年1月, MIT, BAAI): Industry-leading BAAI BGE-M3
- Industry-leading multi-functionality dense+sparse+multi-vector M3: Industry-leading multi-functionality dense+sparse+multi-vector
- Industry-leading 100+ languages multilingual: Industry-leading 100+ languages multilingual
- Industry-leading 8192 max tokens vs 512 typical: Industry-leading 8192 max tokens
- Industry-leading 66.4 BEIR + MTEB top-tier: Industry-leading 66.4 BEIR + MTEB top-tier
- 競合 OpenAI text-embedding-3-small + Cohere embed-v3 + Voyage AI voyage-3 + Nomic Embed Text v1.5: Industry-leading embedding model competitors
選び方・注意点
BGE-M3 BAAI は「Industry-leading multi-functionality + 100+ languages」「Industry-leading 8192 tokens + BAAI Beijing 6 years」用途のIndustry-leading premium 2024年1月 BAAI product。Industry-leading multi-functionality (Industry-leading multi-functionality dense + sparse + multi-vector signature M3 = Multi-Granularity + Multi-Lingual + Multi-Functionality + Industry-leading BGE-M3 multi-functionality signature + Industry-leading hybrid retrieval embedding) で Industry-leading multi-functionality + Industry-leading BGE-M3 + Industry-leading hybrid retrieval embedding。Industry-leading 100+ languages (Industry-leading 100+ languages multilingual signature + Industry-leading BGE-M3 multilingual signature + Industry-leading global language coverage) で Industry-leading 100+ languages + Industry-leading BGE-M3 + Industry-leading global language coverage。Industry-leading 8192 max tokens (Industry-leading 8192 max tokens signature vs 512 typical + Industry-leading BGE-M3 long-context signature + Industry-leading long-document embedding) で Industry-leading 8192 max tokens + Industry-leading BGE-M3 + Industry-leading long-document embedding。Industry-leading sparse+multi-vector (Industry-leading sparse vector lexical matching + Industry-leading multi-vector ColBERT-like signature + Industry-leading BGE-M3 hybrid retrieval modalities + Industry-leading dense+sparse+multi-vector combined) で Industry-leading sparse+multi-vector + Industry-leading BGE-M3 hybrid + Industry-leading dense+sparse+multi-vector combined。Industry-leading BAAI 2018 Beijing (Industry-leading BAAI 2018 Beijing China 6 years + Industry-leading 1024d dense vector + Industry-leading 66.4 BEIR retrieval score + Industry-leading MTEB top-tier + Industry-leading HuggingFace + FlagEmbedding native + Industry-leading MIT license) で Industry-leading BAAI 2018 + Industry-leading Beijing + Industry-leading 6 years + Industry-leading 1024d dense + Industry-leading 66.4 BEIR + Industry-leading MTEB top-tier + Industry-leading HuggingFace+FlagEmbedding + Industry-leading MIT license。但しIndustry-leading OpenAI text-embedding-3-small + Cohere embed-v3 + Voyage AI voyage-3 + Nomic Embed Text v1.5 competition で Industry-leading 4-embedding model competitors vs BGE-M3 BAAI + Industry-leading multi-functionality dense + sparse + multi-vector M3 + 100+ languages multilingual + 8192 max tokens vs 512 typical + 1024d dense vector + sparse vector lexical matching + multi-vector ColBERT-like + 66.4 BEIR retrieval score + MTEB top-tier + HuggingFace + FlagEmbedding native + BAAI 2018 Beijing China 6 years + MIT license unique advantage adoption alignment必須。
関連用語との違い
- vs OpenAI text-embedding-3-small (2024 $0.02/1M): BGE-M3はFree open-source + multi-functionality + 8192 tokens BAAI・text-embedding-3-smallはAPI commercial $0.02/1M + Matryoshka + 8192 tokens OpenAI
- vs Cohere embed-v3 (2024 $0.10/1M): BGE-M3はFree open-source + multi-functionality BAAI Beijing・embed-v3はAPI commercial $0.10/1M + 100+ languages Cohere Toronto
- vs Nomic Embed Text v1.5 (2024 Free Apache 2.0): BGE-M3はMIT + multi-functionality + 8192 tokens BAAI・Nomic Embed Text v1.5はApache 2.0 + Matryoshka 64-768d + 8192 tokens Nomic AI
よくある質問（FAQ）
Q1: BGE-M3 vs Cohere embed-v3 違いは？ A: BGE-M3 (Industry-leading multi-functionality dense+sparse+multi-vector M3 + 100+ languages multilingual + 8192 max tokens + 1024d dense + sparse vector + multi-vector ColBERT-like + 66.4 BEIR + MTEB top-tier + HuggingFace+FlagEmbedding + BAAI 2018 Beijing 6 years + MIT) vs embed-v3 (Industry-leading 1024d + 100+ languages + 64.5 MTEB + 51.4 MIRACL + embed-english-v3.0 + embed-multilingual-v3.0 + embed-light-v3.0 + Cohere API + AWS Bedrock + Azure + input_type + Cohere 2019 Toronto Aidan Gomez 5 years + $0.10/1M)・Industry-leading multi-functionality + dense+sparse+multi-vector + 100+ languages + 8192 tokens + 66.4 BEIR + HuggingFace + BAAI + Beijing + 2018 + 6 years + MIT = BGE-M3 + Industry-leading 1024d + 100+ languages + 64.5 MTEB + 51.4 MIRACL + 3-variant + AWS Bedrock + Azure + input_type + Cohere + Toronto + Aidan Gomez + 2019 + 5 years + $0.10/1M = embed-v3 preference judgment (BGE-M3 open-source multi-functionality + 8192 tokens self-hosted vs embed-v3 API commercial Cohere ecosystem differ)。
Q2: Industry-leading multi-functionality + 100+ languages value は？ A: Industry-leading multi-functionality + 100+ languages (Industry-leading multi-functionality dense + sparse + multi-vector M3 + Industry-leading 100+ languages multilingual + Industry-leading BGE-M3 dual signature)。
Q3: Industry-leading 8192 tokens + BAAI Beijing 6 years value は？ A: Industry-leading 8192 tokens + BAAI (Industry-leading 8192 max tokens vs 512 typical + Industry-leading BAAI Beijing Academy of Artificial Intelligence 2018 Beijing China 6 years heritage)。
まとめ
BGE-M3 BAAI = 2024年1月BAAI発表のpremium MIT open-source multi-functionality embedding model。Industry-leading multi-functionality dense + sparse + multi-vector signature M3 = Multi-Granularity + Multi-Lingual + Multi-Functionality + Industry-leading 100+ languages multilingual signature + Industry-leading 8192 max tokens signature vs 512 typical + Industry-leading 1024d dense vector + Industry-leading sparse vector lexical matching signature + Industry-leading multi-vector ColBERT-like signature + Industry-leading 66.4 BEIR retrieval score + Industry-leading MTEB top-tier signature + Industry-leading HuggingFace + FlagEmbedding native + Industry-leading BAAI 2018 Beijing China 6 years + Industry-leading MIT license + Industry-leading BGE-M3 BAAI MIT open-source multi-functionality embedding 2024 position確立。