BGE-M3

BGE-M3 is BAAI's January 2024 multilingual embedding snapshot, not a 2026 retrieval crown. Rechecked 2026-08-17: huggingface.co/BAAI/bge-m3 is still listed, MIT, XLM-RoBERTa, 568M, 1024-d, 8192 tokens, with leftover dense / multi-vector / sparse copy. Drop leftover first model to do all three / MIRACL nDCG@10 70.0 / MKQA 75.5% / beats OpenAI latest embedding as a 2026 census.

Compare text-embedding-3-large if you wanted a still-listed OpenAI v3 row, and text-embedding-ada-002 if you wanted the older billed OpenAI snapshot.

Key Features

• Previous open card: The HF page we opened still hosts the 2024 weights. Recheck that card before you quote 568M / 1024-d / 8192.
• Not a 2026 leaderboard: MIRACL 70.0 and MKQA 75.5 stay off the table unless you re-run a first-party board.
• Hybrid story is leftover: Dense + multi-vector + sparse is still on the card. Do not call it the first or only hybrid embedder in 2026.
• No invented API dollar: Self-host or recheck a host. Do not invent a BGE-M3 token price.

Use Cases

• People with a 2024 BGE-M3 bookmark who need the catalog corrected.
• People comparing leftover hybrid BGE vs billed OpenAI v3 rows.
• People who were about to paste MIRACL 70.0 into a 2026 deck.

Limitation: We did not re-embed a multilingual corpus or re-run MIRACL / MKQA.

Pricing

First-party pages on 2026-08-17.

If a leftover blog still calls BGE-M3 the current multilingual crown, treat this page as the correction.

Getting Started

1. Open huggingface.co/BAAI/bge-m3 before you download weights.
2. Recheck text-embedding-3-large if you wanted a billed OpenAI v3 row.
3. Recheck text-embedding-3-small if you wanted the cheaper listed v3 row.
4. Do not paste MIRACL 70.0 into a deck.

Frequently Asked Questions

Is BGE-M3 still the multilingual embedding flagship?

Not as a 2026 census. It is a still-listed 2024 MIT card.

Did it beat OpenAI on MKQA?

That leftover score is out as a 2026 census.

Same as text-embedding-3-large?

No. 3-large is a billed OpenAI v3 row. BGE-M3 is an open 2024 snapshot.

Alternatives

• text-embedding-3-large: Still-listed OpenAI v3 large row.
• text-embedding-3-small: Cheaper listed OpenAI v3 row.
• text-embedding-ada-002: Older billed OpenAI snapshot.

Tips

1. Call BGE-M3 a January 2024 snapshot.
2. Recheck the HF card before you quote 8192 tokens or 1024-d.
3. Do not invent a 2026 MTEB rank.

Conclusion

BGE-M3 is a 2024 BAAI multilingual embedding snapshot that is still on Hugging Face, not a 2026 MIRACL or MTEB crown. Start at huggingface.co/BAAI/bge-m3, then decide whether an OpenAI v3 row already covers the retrieval path you need.