Embeddings BGE-M3 pour un RAG sur de longs documents
BGE-M3 pour un RAG sur de longs documents : 1024 dimensions, jusqu'à 8192 tokens, plus de 100 langues, sans préfixe. Licence MIT.
| Model | BAAI/bge-m3 |
|---|---|
| Dimensions | 1024 |
| Max tokens | 8.2K |
| Distance | cosine |
| Languages | fr, en, de, es, it, nl, pt, zh, ja |
| Chunk size | 1000 |
Notes
Pas de préfixe à ajouter devant les questions ni les passages. Accepte des textes longs (8192 tokens) : on peut découper en gros morceaux (environ 1000 mots). Le modèle sait aussi produire des vecteurs creux (recherche par mots-clés) avec la bibliothèque FlagEmbedding. Modèle publié par BAAI sous licence MIT.
Estimated cost
Input tokens per call : ≈ 239
| Model | Per call | Per 1,000 calls |
|---|---|---|
| Mistral Large · Mistral AI | €0.00011 | €0.11 |
| DeepSeek Flash · DeepSeek | €0.00006 | €0.06 |
| DeepSeek V4 Pro · DeepSeek | €0.00028 | €0.28 |
| GPT-6 Luna · OpenAI | €0.00002 | €0.02 |
| GPT-6 Sol · OpenAI | €0.00042 | €0.42 |
| Claude Sonnet 5.5 · Anthropic | €0.00042 | €0.42 |
| Claude Opus 5.5 · Anthropic | €0.00084 | €0.84 |
| Self-hosted open-source model (Ollama, vLLM) | €0 in API fees (server cost only) | |
Input cost only, excluding the model's answer. Providers' public standard prices (Mistral AI, DeepSeek, OpenAI, Anthropic) converted from USD to euros at the ECB rate of September 29, 2026. Estimate: 1 token ≈ 3.6 characters.
Common to all 3 regions
Sources under a compatible licence.
Sources : Modèle BAAI/bge-m3, licence MIT (fiche Hugging Face)
Commercial use allowed · credit the author · changes allowed.
Text or data only: no access to files, network or commands.
Hosted in France (Scaleway, Paris). No transfer outside the EU.
European Union · 5/5
No personal data.
No training data: no summary required.
United States · 5/5
No personal data.
No training data: no documentation to publish.
China · 5/5
No personal data.
No training data; AI-generated content published in China must be labelled (2025).
Indicative summary as of 30/09/2026, not a legal certification. Method and sources →
Statistics
Likes : 0 · Comments : 0
Click a counter to show or hide it, hover the chart for details. One view per visitor per day, bots excluded; counting started on 30 September 2026.
- Identifier
- contextetech--bge-m3
- Type
- Embeddings
- File
- bge-m3.embedding.json
- Format
- JSON (application/json)
- Size
- 636 bytes
- Encoding
- UTF-8
- Estimated tokens
- ≈ 176
- Language
- French
- License
- MIT
- Commercial use
- Allowed
- Personal data
- None
- Version
- 1.0
- Published on
- October 1, 2026 at 6:21 PM
- Updated on
- October 1, 2026 at 6:21 PM
- Storage
- in database
- Uses
- 0
- Likes
- 0
bge-m3.embedding.json
{
"format": "contextetech/embedding/v1",
"id": "contextetech/bge-m3",
"description": "BGE-M3 pour un RAG sur de longs documents : 1024 dimensions, jusqu'à 8192 tokens, plus de 100 langues, sans préfixe. Licence MIT.",
"tags": [
"rag",
"multilingue",
"longs-documents"
],
"license": "MIT",
"language": "fr",
"model": "BAAI/bge-m3",
"dimensions": 1024,
"maxTokens": 8192,
"distance": "cosine",
"languages": [
"fr",
"en",
"de",
"es",
"it",
"nl",
"pt",
"zh",
"ja"
],
"prefixes": {
"query": "",
"passage": ""
},
"chunkSize": 1000,
"benchmarks": []
}Embed (Python)
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("BAAI/bge-m3")
q = model.encode(["your input here"], normalize_embeddings=True)
docs = model.encode(["…"], normalize_embeddings=True)
print(q.shape) # (1, 1024)
print(q @ docs.T) # similarité
Community
No comments yet. Share your feedback, it will help the next person.