KaLM Embedding Multilingual Mini Instruct V1
HIT-TMG/KaLM-embedding-multilingual-mini-instruct-v1
published Oct 2024 · updated Mar 2025
KaLM Embedding Multilingual Mini Instruct V1 is a multilingual embedding model adapted from Qwen2-0.5B that supports instruction-based queries for tasks like retrieval, classification, and clustering.
specs
| Task | Embedding |
| Architecture | Adapted auto-regressive LLM (Qwen2-0.5B) with mean pooling |
| Parameters | 494M |
| License | MIT |
| Max Sequence Length | 512 tokens |
| Instruction Support | Yes (for classification and clustering) |
best for
- ·Multilingual text retrieval and semantic search
- ·Text classification with instruction-based prompts
- ·Clustering of multilingual sentences
FAQ
It is best for multilingual embedding tasks such as retrieval, classification, and clustering, with optional instruction prompts for classification and clustering.
It outperforms similar-sized models like multilingual-e5-large and bge-m3 on the MTEB benchmark, achieving an average score of 64.16.
Input is text strings; output is normalized embedding vectors (dense). The model supports a prompt prefix for instruction tasks and a max sequence length of 512 tokens.
Use the gigarouter OpenAI-compatible endpoint with your API key. Send a request with the model name and input text to get the embedding vector.
The KaLM-Embedding repository is released under the MIT license, allowing free use, modification, and distribution.
We're benchmarking and onboarding KaLM Embedding Multilingual Mini Instruct V1 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.