Skip to content
← Models
Open weightsBAAI/10 trustVerified
BGE-M3 logo

BGE-M3

The world's most-downloaded embedding model — 32.8M downloads of multilingual, multi-function retrieval.

Visit official page →

About

BGE-M3 (June 2024, Beijing Academy of AI) is the most-downloaded embedding model on Hugging Face (32.8M downloads): multilingual (100+ languages), multi-function (dense + sparse + ColBERT multi-vector in one model), and 8K context.

MIT licensed — commercial use fully allowed. The multi-function design means one checkpoint serves dense retrieval, lexical (sparse) matching, and late-interaction reranking — a genuine architectural differentiator. It became the default embedding for multilingual RAG worldwide. For English-only workloads the smaller BGE-large-en-v1.5 (also in this batch) is leaner; for maximum multilingual quality this remains a top answer.

Specs

Context window

8,192 tokens

Pricing

Free — open weights (MIT), self-hosted, commercial use allowed.

Free quota

Free — download and run the weights yourself (MIT).

API access

Self-hosted (sentence-transformers, FlagEmbedding; Hugging Face checkpoints).

Weights

Open weights

Released

June 4, 2024

Benchmarks

Provider-claimed (BAAI): leading multilingual MTEB performance at release — treat as provider-claimed. Independent: 32.8M downloads is the strongest community validation in embeddings; widely third-party benchmarked with strong results.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

ragsearchmultilingualsemantic-searchreranking

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier). The numbers speak: 32.8M downloads (the new HF peer max), MIT license, multi-function architecture, and two years of community validation as the multilingual RAG default. Docked two points: (1) the 2024 release means maintenance score is dated; (2) newer architectures are emerging. The reference open embedding model.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
10.0/10
Maintenance · recency, releases, issue load
4.0/10
Openness · license + docs
9.7/10
Manual review · sanity read + red-flag check
10.0/10
  • · HF downloads 32,815,892 (peer-normalized)
  • · released 861 days ago
  • · manual review: the reference open embedding model; multi-function architecture; two years of community validation; license: MIT

Sources