BGE-M4 embeddings, one SDK call
The successor to the open embedding model that made multilingual retrieval boring. One vector space, one endpoint, per-token pricing, through the same OpenAI SDK call you already use.
- Open weights expected
- Multilingual retrieval
- Input-only metering
- Per token, no contract
jk. It is not out yet.
BAAI has not announced BGE-M4, and nothing credible is circulating about one. BGE-M3 is still the flagship, it is live here now, and BAAI's own answer for better retrieval today is not a new embedder at all.
Below is every rumor we could source, what BAAI has actually confirmed, and the model you can call in the meantime. When the real one ships, this page becomes its page at this URL.
The BGE-M4 rumor mill.
Each row labelled by how much it is worth. Last checked September 24, 2026.
- Release status
- ConfirmedUnreleased. BGE-M3 remains BAAI's flagship multilingual embedder. The M3 is multi-linguality (100+ languages), multi-granularity (inputs up to 8192 tokens), and multi-functionality (dense, sparse, and multi-vector retrieval from one model). BAAI/bge-m3
- Leaks
- No signalNothing credible names a BGE-M4. No leak, no paper, no model card.
- BAAI's own answer
- ConfirmedThe current recommendation for better retrieval is a pipeline change rather than a new embedder: hybrid retrieval plus reranking, with bge-reranker-v2-m3 built on M3. If you are waiting on a new model to fix recall, that is the thing to try first. FlagEmbedding
- License
- UnverifiedOpen weights on Hugging Face expected, the way the whole BGE family has shipped.
- How it would serve here
- ConfirmedSelf-hosted on our own GPUs behind POST /v1/embeddings, the same surface BGE-M3 uses today, with input-token metering only.
- Price
- No signalUnknown until we have the model. Embeddings meter input tokens only, as they do today.
What you can call right now.
The BAAI models live on the API today, at the rates they bill at.
Full inference catalogWhy it shows up here early.
Not a promise about a date. A description of the path a new open model takes to get behind this API.
The embeddings surface is already here
BGE-M3 serves on our own GPUs through the standard embeddings endpoint, including asynchronous batch for full index builds. A new embedding model drops into that rather than needing it built.
Open weights make this a days problem
Embedding models are small next to the chat catalog, so the gap between a public weights drop and a served endpoint is short.
Your index stays portable either way
Open-weight embeddings mean nothing proprietary is baked into your vectors. When you decide to re-embed on a newer model, that is your call on your schedule, not a migration forced by a vendor.
One email, the day it is callable.
Leave an address and we send exactly one message when BGE-M4 is serving here, with the model id and the rate. Nothing else goes to this list.
BGE-M4, answered.
Is BGE-M4 out?
No. As of September 2026 BAAI has not announced a successor to BGE-M3, and nothing credible is circulating about one. BGE-M3 is current and it runs on OpenRelay today.
Should I wait to build my index?
No, and this is the answer even though it costs us the signup. Embedding a corpus on BGE-M3 costs single-digit dollars at catalog rates for most collections, and re-embedding later costs the same again. Waiting on an unannounced model is the more expensive choice.
My retrieval is not good enough. Is a new model the fix?
Usually not. BAAI's own recommendation is hybrid retrieval plus reranking, using bge-reranker-v2-m3 on top of M3, which is a pipeline change you can make today rather than a model you are waiting for.
When is the BGE-M4 release date?
Unannounced. BAAI publishes BGE weights to Hugging Face; when a successor lands this page becomes the live model page and the notify list gets an email.
Will OpenRelay serve the next BGE model?
Assuming open weights, yes. We self-host BGE-M3 on our own GPUs behind the embeddings endpoint, and adding a sibling model to that surface is small work.
What should I use until then?
BGE-M3, on the card above. One SDK call, 100+ languages in one vector space, 8K-token inputs, and the Batch API for the initial index build at half the realtime input rate.
More models
The account outlasts the model.
One API key, one base URL, per-token billing. Whatever BAAI ships next, adopting it is a one-line change rather than a new vendor.