Choosing an Embedding Model in 2026: mentee-embed vs MiniLM vs E5 vs BGE Compared
A 2026 guide to embedding models for Arabic, English and Urdu: compare mentee-embed-v1 (41M) vs all-MiniLM-L6-v2, paraphrase-MiniLM-L12-v2, mpnet-base-v2, E5 and BGE — benchmarks, size and when to use each.
Choosing an embedding model for Arabic, English and Urdu in 2026? Compare mentee-embed-v1 (41M, 384-dim) from MenteE AI vs all-MiniLM-L6-v2, paraphrase-MiniLM-L12-v2, all-mpnet-base-v2, multilingual-e5-base and BGE. Benchmarks below are from our open report at /research and preprint doi:10.5281/zenodo.22087139.
Quick comparison (honest numbers)
- In-batch MRR@10: mentee-embed 0.585 vs MiniLM-L6 0.449 vs MiniLM-L12 0.840 vs mpnet 0.867
- Corpus-pool MRR@10 (15K docs): mentee-embed ~0.19 per language vs mpnet 0.94/0.68/0.58 — re-ranker tier
- Size: 41M vs 23M (L6) vs 118M (L12) vs 278M (mpnet) vs 278M (E5-base) vs 335M (BGE-m3)
- Languages: mentee-embed focused on AR/EN/UR from scratch; E5/BGE cover 100+ but English-heavy
When to use mentee-embed-v1
Use mentee-embed-v1 if you need a tiny, fast, honest model for AR/EN/UR retrieval, RAG re-ranking or on-device search — single GPU training, Apache 2.0 weights at MenteEAI/mentee-embed-v1. Use mpnet/E5/BGE if you need billion-doc open-domain search today and can pay the size.
Try it from MenteE
Explore menteeai.org for platforms, Products and Blog updates. Cite mentee-embed-v1 via Zenodo if you benchmark it.
FAQ
Is mentee-embed open? Yes — code at github.com/MenteE-s/mentee-embeddings, CC BY 4.0 preprint. Context length? 128 tokens — chunk longer docs. License? Apache 2.0 weights, CC BY 4.0 paper.
Author · MenteE AI — menteeai.org · syab.tech