BAAI/bge-m3 embeddings API: 100+ languages, $0.02 per 1M
BAAI/bge-m3 is available on Lazu through the OpenAI-compatible /v1/embeddings endpoint at $0.02 per million input tokens. It returns 1024-dimension vectors, accepts up to 8,192 tokens per input and puts more than 100 languages into one vector space. Embedding 10,000 documents of 500 tokens each costs about $0.10.
What bge-m3 is #
bge-m3 is an open-weights embedding model from BAAI, the Beijing Academy of Artificial Intelligence, published in January 2024 under the MIT license. "M3" stands for multi-linguality (100+ languages), multi-granularity (from a short sentence to an 8,192-token document) and multi-functionality (dense, sparse and multi-vector retrieval). Lazu's /v1/embeddings returns the dense vector, the form vector databases and RAG frameworks expect. Details are in the model card and the paper.
At a glance #
BAAI/bge-m3 on Lazu |
|
|---|---|
| Endpoint | POST /v1/embeddings, OpenAI-compatible |
| Price | $0.02 per million input tokens; embeddings have no output tokens |
| Vector size | 1024, fixed |
| Max tokens per input | 8,192 |
| Languages | 100+, one shared vector space |
dimensions, encoding_format: base64 |
Ignored: always 1024 floats |
| Batch | An array of inputs per request (we tested 50) |
Call it #
Any OpenAI SDK works: point base_url at Lazu and use the model ID BAAI/bge-m3.
curl https://api.lazu.ai/v1/embeddings \
-H "Authorization: Bearer $LAZU_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "BAAI/bge-m3", "input": ["How do I reset my password?", "如何重置密码?"]}'
from openai import OpenAI
client = OpenAI(base_url="https://api.lazu.ai/v1", api_key="YOUR_LAZU_KEY")
resp = client.embeddings.create(
model="BAAI/bge-m3",
input=["How do I reset my password?", "如何重置密码?"],
)
print(len(resp.data[0].embedding), resp.usage.prompt_tokens, resp.usage.cost)
The response has one vector per input in input order, model set to BAAI/bge-m3, and usage with prompt_tokens, total_tokens and cost in USD.
Search across languages #
Because every language lands in the same space, an English question finds a Chinese or Japanese answer without translating anything:
import numpy as np
docs = [
"Reset your password from Settings → Security.",
"账单每月 1 日生成,可在控制台下载发票。",
"API キーはプロジェクトごとに発行できます。",
]
def embed(texts):
resp = client.embeddings.create(model="BAAI/bge-m3", input=texts)
vectors = np.array([item.embedding for item in resp.data])
return vectors / np.linalg.norm(vectors, axis=1, keepdims=True)
scores = embed(docs) @ embed(["Where can I download my invoice?"])[0]
print(docs[int(scores.argmax())]) # the Chinese invoice sentence
Limits and errors #
- An input over 8,192 tokens is refused with
400 context_length_exceeded(LZ-1005) and costs nothing. Chunk long documents first; chunks of a few hundred to a thousand tokens usually retrieve better anyway. - An empty string is refused with a 400. Drop empty chunks before you send them.
dimensionsandencoding_format: base64are ignored. You always get 1024 floats. The catalog lists both underunsupported_paramsfor this model.
What you are billed #
Only input tokens are billed, and failed requests cost nothing. The provider behind this model does not report a token count, so Lazu counts the input with a standard tokenizer. That count differs from bge-m3's own tokenizer by roughly a fifth to a quarter: lower for English text, higher for Chinese. At $0.02 per million tokens, that gap is worth at most about $0.005 per million tokens embedded.
Every response carries the charge in usage.cost and in the x-lazu-cost-usd header. What we measured on October 29, 2026:
| Request | Billed tokens | Charge |
|---|---|---|
| One short English sentence | 7 | $0.000001 (minimum per request) |
| English text, 10,500 characters | 1,948 | $0.000039 |
| Chinese text, 2,880 characters | 3,053 | $0.000061 |
| Batch of 20 short documents | 260 | $0.000005 |
bge-m3 or text-embedding-3-small? #
BAAI/bge-m3 |
text-embedding-3-small |
|
|---|---|---|
| Price on Lazu | $0.02 / 1M tokens | $0.02 / 1M tokens |
| Vector size | 1024 | 1536, can be shortened |
| Max tokens per input | 8,192 | 8,191 |
| Weights | Open, MIT license | Closed |
Pick bge-m3 for Chinese, Japanese or mixed-language content, or when a smaller index matters: 1024 dimensions take a third less storage than 1536. Keep text-embedding-3-small if your index already holds its vectors. Vectors from different models cannot be compared, so switching models means re-embedding the whole corpus.
The full request and response reference is in the embeddings docs.