--- title: BAAI/bge-m3 (Encode) description: "For more details please refer to our github repo: https://github.com/FlagOpen/FlagEmbedding. XLM-RoBERTa, 568M parameters." canonical_url: https://superlinked.com/models/baai-bge-m3 last_updated: 2026-04-22 --- # BAAI/bge-m3 (Encode) For more details please refer to our github repo: https://github.com/FlagOpen/FlagEmbedding Source: [BAAI/bge-m3 on HuggingFace](https://huggingface.co/BAAI/bge-m3) ## Overview | Field | Value | |-------|-------| | Architecture | XLM-RoBERTa | | Parameters | 568M | | Tasks | Encode | | Outputs | Dense, Sparse, Multi-Vec | | Dimensions | Dense: 1,024, Sparse: 250,002, Multi-Vec: 1,024 | | Max sequence length | 8,192 tokens | | License | mit | | Inputs | text | ## Benchmarks ### CQADupstackPhysicsRetrieval Domain: scientific · Task: retrieval · Language: en Duplicate question retrieval from StackExchange Physics Corpus: 38,314 · Queries: 1,039 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 28.0K tok/s · Corpus p50 70.7ms · Query 3.2K tok/s · Query p50 44.7ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 27.3K tok/s · Corpus p50 74.8ms · Query 3.1K tok/s · Query p50 45.2ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 27.8K tok/s · Corpus p50 70.8ms · Query 3.3K tok/s · Query p50 44.9ms [Reference](http://nlp.cis.unimelb.edu.au/resources/cqadupstack/) ### CosQA Domain: technology · Task: retrieval · Language: en Code search with natural language queries Corpus: 6,267 · Queries: 500 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 15.3K tok/s · Corpus p50 51.4ms · Query 1.8K tok/s · Query p50 45.8ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 15.8K tok/s · Corpus p50 51.3ms · Query 1.7K tok/s · Query p50 47.9ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 16.9K tok/s · Corpus p50 51.0ms · Query 1.9K tok/s · Query p50 43.7ms [Reference](https://arxiv.org/abs/2105.13239) ### FiQA2018 Domain: finance · Task: retrieval · Language: en Financial opinion mining and question answering Corpus: 57,599 · Queries: 648 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 33.8K tok/s · Corpus p50 77.8ms · Query 3.6K tok/s · Query p50 44.8ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 28.6K tok/s · Corpus p50 87.3ms · Query 3.1K tok/s · Query p50 48.4ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 30.2K tok/s · Corpus p50 81.3ms · Query 3.3K tok/s · Query p50 46.2ms [Reference](https://sites.google.com/view/fiqa/) ### LegalBenchConsumerContractsQA Domain: legal · Task: retrieval · Language: en Question answering on consumer contracts Corpus: 153 · Queries: 396 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 39.1K tok/s · Corpus p50 216.2ms · Query 5.1K tok/s · Query p50 44.0ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 39.5K tok/s · Corpus p50 210.7ms · Query 4.2K tok/s · Query p50 51.7ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 40.6K tok/s · Corpus p50 208.1ms · Query 5.1K tok/s · Query p50 45.0ms [Reference](https://huggingface.co/datasets/nguha/legalbench) ### NFCorpus Domain: medical · Task: retrieval · Language: en Biomedical literature search from NutritionFacts.org Corpus: 3,593 · Queries: 323 **Variant: default** **Quality:** ndcg at 10: 0.3144 · map at 10: 0.1174 · mrr at 10: 0.5243 **Performance (A10G b1 c4):** Corpus 127 tok/s · Corpus p50 9.4s · Query 39 tok/s · Query p50 537.6ms **Performance (L4 b1 c16):** Corpus 38.1K tok/s · Corpus p50 100.0ms · Query 1.5K tok/s · Query p50 41.0ms **Variant: multivector** **Performance (L4 b1 c16):** Corpus 37.9K tok/s · Corpus p50 126.6ms · Query 1.5K tok/s · Query p50 46.1ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 34.2K tok/s · Corpus p50 134.0ms · Query 937 tok/s · Query p50 64.2ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 40.0K tok/s · Corpus p50 124.4ms · Query 1.4K tok/s · Query p50 45.0ms [Reference](https://www.cl.uni-heidelberg.de/statnlpgroup/nfcorpus/) ### NanoFiQA2018Retrieval Domain: finance · Task: retrieval · Language: en Smaller subset of the FiQA financial QA dataset **Variant: multivector** **Performance (L4 b1 c16):** Corpus 30.6K tok/s · Corpus p50 86.6ms · Query 2.7K tok/s · Query p50 54.1ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 28.1K tok/s · Corpus p50 89.9ms · Query 2.5K tok/s · Query p50 59.2ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 30.3K tok/s · Corpus p50 87.5ms · Query 2.7K tok/s · Query p50 53.3ms **Variant: default** **Quality:** ndcg at 10: 0.5726 · map at 10: 0.4957 · mrr at 10: 0.6467 **Performance (L4 b1 c16):** Corpus 31.2K tok/s · Corpus p50 68.9ms · Query 2.9K tok/s · Query p50 43.3ms [Reference](https://sites.google.com/view/fiqa/) ### SCIDOCS Domain: scientific · Task: retrieval · Language: en Citation prediction, document classification, and recommendation for scientific papers Corpus: 25,656 · Queries: 1,000 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 34.4K tok/s · Corpus p50 90.7ms · Query 3.4K tok/s · Query p50 43.0ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 33.2K tok/s · Corpus p50 94.8ms · Query 2.5K tok/s · Query p50 54.2ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 32.5K tok/s · Corpus p50 93.4ms · Query 3.4K tok/s · Query p50 44.7ms [Reference](https://allenai.org/data/scidocs) ### SciFact Domain: scientific · Task: retrieval · Language: en Scientific claim verification using research literature Corpus: 5,183 · Queries: 300 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 38.2K tok/s · Corpus p50 116.5ms · Query 4.9K tok/s · Query p50 44.4ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 37.9K tok/s · Corpus p50 118.2ms · Query 4.2K tok/s · Query p50 48.5ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 37.8K tok/s · Corpus p50 117.6ms · Query 4.2K tok/s · Query p50 47.7ms [Reference](https://github.com/allenai/scifact) ### StackOverflowQA Domain: technology · Task: retrieval · Language: en Programming question answering from Stack Overflow Corpus: 19,931 · Queries: 1,994 **Variant: multivector** **Performance (L4 b1 c16):** Corpus 33.2K tok/s · Corpus p50 107.6ms · Query 33.8K tok/s · Query p50 137.6ms **Variant: dense** **Performance (L4 b1 c16):** Corpus 33.4K tok/s · Corpus p50 104.7ms · Query 33.8K tok/s · Query p50 138.5ms **Variant: sparse** **Performance (L4 b1 c16):** Corpus 32.7K tok/s · Corpus p50 108.4ms · Query 34.8K tok/s · Query p50 137.6ms [Reference](https://arxiv.org/abs/2407.02883)