Baai · since Januar/Februar 2024 (GitHub-Release 30.01.2024; arXiv-Paper veröffentlicht 5. Februar 2024) · 2× · last seen Jun 30, 2026
BGE-M3 is an open-source embedding model by the Beijing Academy of Artificial Intelligence (BAAI), based on a fine-tuned XLM-RoBERTa architecture that encodes text into 1024-dimensional vectors. It is distinguished by three properties: support for more than 100 languages (Multi-Linguality), simultaneous dense, sparse, and multi-vector retrieval (Multi-Functionality), and processing of inputs from short sentences up to 8192 tokens (Multi-Granularity). The model is freely available via Hugging Face, GitHub (FlagEmbedding) and various cloud platforms such as NVIDIA NIM or Ollama, and is distributed under the MIT license.
Features
| Deployment (Self-host/Cloud) | Self-hosting via Python-Bibliothek (FlagEmbedding, Transformers, ONNX) oder Cloud-Inferenz via NVIDIA NIM/Ollama |
| Durchsatz/Latenz | Keine offiziellen Benchmark-Zahlen zu Durchsatz/Latenz in den geprüften Quellen gefunden |
| Lizenz | MIT-Lizenz, frei für akademische und kommerzielle Nutzung |
| Plattform | Hugging Face, GitHub (FlagEmbedding), NVIDIA NIM, Ollama, ONNX |
| Preis | Kostenlos, Open-Source-Gewichte (Download via Hugging Face/GitHub) |
| Protokoll-Kompatibilität | Max. 8192 Tokens Kontextlänge; unterstützt Dense-, Sparse- (BM25-ähnlich) und Multi-Vektor-Retrieval (ColBERT-Stil); Integration mit Milvus und Vespa für Hybrid-Retrieval |
| Release-Datum | 30. Januar 2024 (GitHub-Ankündigung); Paper veröffentlicht 5. Februar 2024 |
| Unterstützte Modelle/Provider | Basis: XLM-RoBERTa (fine-tuned), 568 Mio. Parameter, Embedding-Dimension 1024; nutzbar via FlagEmbedding, sentence-transformers, HuggingFace transformers, LangChain, Milvus, Vespa |