Back to Gap Map Model components / Embeddings & retrieval models

BGE

Beijing Academy of Artificial Intelligence (BAAI)

open weights / Overall score: 4.3(strong)

BGE (BAAI General Embedding) covers text embedding and reranking for retrieval pipelines. The bge-m3 checkpoint works across more than 100 languages and inputs up to 8,192 tokens, emitting dense, sparse and multi-vector representations from a single pass; the English v1.5 encoders span small to large, and the bge-reranker line supplies cross-encoder relevance scores. The Beijing Academy of Artificial Intelligence maintains the family and the FlagEmbedding toolkit around it.

The family ships under three different licenses, and the two Gemma-based checkpoints are the ones that govern the openness score.

Openness

3 high confidence

3.0

weights: open(ungated safetensors and PyTorch checkpoints across the v1.5 encoders and bge-m3)
data: open(the BAAI-MTP pretraining corpus, the bge-m3-data fine-tuning set and the MLDR long-document set are all published)
code: open(FlagEmbedding ships embedder and reranker fine-tuning, unified fine-tuning, evaluation and inference pipelines under MIT)
license: MIT(OSI

The training data and code pipeline are fully published, so the license is what limits the score. Most of the family ships under MIT or Apache-2.0, but bge-multilingual-gemma2 carries the more restrictive Gemma license from Gemma 2, and that shipped checkpoint sets the score.

https://huggingface.co/api/models/BAAI/bge-m3 recorded 2026-09-11
license: mit; gated: false; private: false; pytorch_model.bin and an ONNX export present.
https://huggingface.co/api/models/BAAI/bge-multilingual-gemma2 recorded 2026-09-11
license: gemma; gated: false; private: false; a four-shard safetensors checkpoint is distributed under the Gemma terms.
https://huggingface.co/api/models/BAAI/bge-reranker-v2-m3 recorded 2026-09-11
license: apache-2.0; gated: false; private: false; model.safetensors present.
https://huggingface.co/api/models/BAAI/bge-small-en-v1.5 recorded 2026-09-11
license: mit; gated: false; private: false; model.safetensors, pytorch_model.bin and ONNX present.
https://huggingface.co/BAAI/bge-m3/raw/main/README.md recorded 2026-09-11
Links the released bge-m3-data fine-tuning set and the MLDR long-document dataset, the unified fine-tuning example, and the MKQA and MLDR evaluation scripts.
https://raw.githubusercontent.com/FlagOpen/FlagEmbedding/master/README.md recorded 2026-09-11
Records that "the massive training data of BGE has been released" at data.baai.ac.cn, links embedder and reranker fine-tuning pipelines and a dataset directory, and states the toolkit is MIT-licensed.

Adoption

5 high confidence

5.0

Adoption is measured as Hugging Face downloads across the BGE embedding checkpoints, led by bge-small-en-v1.5 and bge-m3. The reranker checkpoints are excluded because they are scored separately as bge-reranker.

https://huggingface.co/api/models/BAAI/bge-m3 recorded 2026-09-11
downloads: 37891167.
https://huggingface.co/api/models/BAAI/bge-small-en-v1.5 recorded 2026-09-11
downloads: 64850135.

Capability

4 high confidence

4.0

Competitive frontier rather than leader: multilingual, long-context, and the best MIRACL figure in the comparison that ranks it, but its BEIR average trails the newer encoders and the family has not shipped a successor to m3.

https://huggingface.co/BAAI/bge-m3/raw/main/README.md recorded 2026-09-11
States support for more than 100 working languages and inputs up to 8192 tokens, with dense, sparse and ColBERT-style multi-vector retrieval from one model.
https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0/raw/main/README.md recorded 2026-09-11
Comparison table gives bge-m3 (BAAI) BEIR (15) 48.8, MIRACL (4) 56.8, CLEF (Focused) 40.8 and CLEF (Full) 41.3, all average NDCG@10.

Verified 2026-09-11