BAAI BGE
Open BGE embedding and reranker models from the Beijing Academy of AI: bge-m3 with 8,192 tokens, 100+ languages, MIT license
Overview
BGE (BAAI General Embedding) is a family of open embedding and reranker models from the Beijing Academy of Artificial Intelligence, a non-profit research lab, developed in the FlagEmbedding repo. The flagship, bge-m3, gives 1,024-dimension vectors for inputs up to 8,192 tokens in more than 100 languages, and does dense, sparse (lexical) and multi-vector retrieval from one model. It is MIT licensed and had over 29 million Hugging Face downloads in the last month. Smaller English models (bge-small-en-v1.5 at 33.4M parameters and 384 dims, bge-large-en-v1.5 at 0.3B and 1,024 dims, both 512 tokens) remain heavily used, and bge-reranker-v2-m3 (0.6B, Apache 2.0) handles reranking. Larger LLM-based options (bge-en-icl at 7B, bge-multilingual-gemma2 under the Gemma license) and the BGE-VL image models round out the line. The newest release is BGE-VL (March 2025); there has been no new text embedding model since 2024. BAAI was added to the US Entity List in March 2025, which some US buyers will need to review.
The Verdict
Who Should Use BAAI BGE?
Best For
- Free commercial self-hosting under MIT
- Multilingual RAG and hybrid (dense plus sparse) search with one model
- Very cheap hosted embeddings through providers such as Cloudflare Workers AI
- Small, fast English embeddings on CPU (bge-small)
Not Ideal For
- Teams that want a vendor API, support or SLA from BAAI
- Organizations whose compliance rules bar Entity List parties
- Inputs longer than 8,192 tokens, or image-heavy search
What's Great
- MIT license on bge-m3 and the v1.5 English models
- Dense, sparse and multi-vector retrieval in one model
- 8,192-token context and 100+ languages
- Huge adoption: tens of millions of monthly downloads
- bge-m3 costs $0.012 per 1M tokens on Cloudflare Workers AI
Watch Out For
- No new text model since 2024; newer models from Voyage, Cohere and Jina post higher vendor scores
- v1.5 English models stop at 512 tokens
- Mixed licenses: bge-multilingual-gemma2 follows Gemma terms, not MIT
- BAAI is on the US Entity List (Federal Register, March 2025)
Pricing
View all features & details
Models
- bge-m3: 1,024 dims, 8,192 tokens, 100+ languages, MIT
- bge-large-en-v1.5: 0.3B, 1,024 dims, 512 tokens, MIT
- bge-small-en-v1.5: 33.4M, 384 dims, 512 tokens, MIT
- bge-reranker-v2-m3: 0.6B, multilingual, Apache 2.0
- bge-en-icl (7B, English) and bge-multilingual-gemma2 (Gemma license)
- BGE-VL: image and text retrieval, MIT
Tooling
- FlagEmbedding library for inference and fine-tuning (MIT)
- Docs at bge-model.com
- Repo still active (pushed October 2026)
Use Cases
- Hybrid search in vector DBs that store sparse and dense vectors
- Multilingual RAG
- Reranking retrieved passages
Real-World Usage
Downloads (Hugging Face, last month)
- bge-small-en-v1.5: about 46.7 million
- bge-m3: about 29.1 million
- bge-reranker-v2-m3: about 14.7 million
- bge-large-en-v1.5: about 8.0 million
Hosting
- Cloudflare Workers AI hosts bge-small, base, large, m3 and a reranker
- 182 models on the BAAI Hugging Face org
- Community advice in 2026 often names bge-m3 the local pick for multilingual RAG
How It Compares
| Feature | BAAI bge-m3 | Nomic Embed v1.5 | Jina v5 text | OpenAI 3-large |
|---|---|---|---|---|
| License | MIT | Apache 2.0 | Check card (most Jina weights CC BY-NC) | Closed API |
| Dimensions | 1,024 | 768 (Matryoshka) | Matryoshka | Up to 3,072 |
| Context | 8,192 tokens | 8,192 tokens | 8K to 32K tokens | 8,192 tokens |
| Languages | 100+ | English | Multilingual | Multilingual |
| Dense, sparse and multi-vector in one model | Yes | Dense | Dense; v4 adds multi-vector | Dense |
| Hosted price per 1M | $0.012 (Cloudflare) | Not published | Top-ups | $0.13 |
Sources: bge-m3 card · Nomic card · Jina embeddings · OpenAI pricing
Is this your tool? Add a badge