MiniCPM-Embedding
openbmb/MiniCPM-Embedding
published Sep 2024 · updated Jan 2025
MiniCPM-Embedding is a bilingual text embedding model that produces dense vector representations for Chinese and English, with strong cross-lingual retrieval capabilities.
specs
| Task | Text Embedding |
| Architecture | Bidirectional attention with Weighted Mean Pooling, based on MiniCPM-2B |
| Parameters | 2.4B |
| Embedding Dimension | 2304 |
| Max Input Tokens | 512 |
| License | Apache-2.0 (code); MiniCPM Model License (weights, free for academic and commercial use after registration) |
best for
- ·Chinese-English cross-lingual document retrieval and search
- ·Semantic search and RAG pipelines for bilingual content
- ·Embedding generation for classification or clustering of Chinese and English text
FAQ
The embedding dimension is 2304 and the maximum input token length is 512.
It supports an optional query-side instruction in the format "Instruction: {{ instruction }} Query: {{ query }}", or instruction-free mode as "Query: {{ query }}". Documents are input directly.
MiniCPM-Embedding has 2.4B parameters. It achieves 76.76 NDCG@10 on C-MTEB/Retrieval and 58.56 on BEIR, outperforming many larger models in cross-lingual tasks.
The code is Apache-2.0. The model weights require following the MiniCPM Model License; they are free for academic research and free for commercial use after filling out a registration questionnaire.
Use the gigarouter OpenAI-compatible endpoint with your API key, sending your text as input to the embeddings endpoint.
We're benchmarking and onboarding MiniCPM-Embedding as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.