2026 Chinese embedding benchmark — bge-m3, Conan, m3e, bce, OpenAI

Setup#

• Corpus: 3,000 FAQ entries + 500 product docs (e-commerce)
• Test set: 120 real user questions with annotated top-1 ground truth
• Retrieval: pure vector, top_k=5
• Metrics: MRR@5, Recall@5, per-query latency, monthly cost per 1M tokens

Models#

Model	Dims	Open	Notes
bge-m3	1024	✓	BAAI; de-facto baseline for Chinese
Conan-embedding-v1	1792	✓	Tencent, 2026
m3e-large	1024	✓	moka-ai
bce-embedding-base_v1	768	✓	NetEase Youdao
text-embedding-3-large	3072	✗	OpenAI

Results#

Model	MRR@5	Recall@5	Avg latency (ms)	Cost / 1M tokens
bge-m3	0.87	0.94	22	self-host GPU
Conan-embedding-v1	0.89	0.95	31	self-host GPU
m3e-large	0.81	0.89	24	self-host GPU
bce-embedding-base_v1	0.79	0.87	18	self-host GPU
text-embedding-3-large	0.85	0.93	180-400	~$0.13

Recommendations#

Scenario	Pick
Default Chinese	bge-m3 — most balanced
Max accuracy	Conan-embedding-v1 — edges out bge-m3
Tight budget	bce-embedding-base_v1 — light and fast
Mixed languages	bge-m3 or text-embedding-3-large
No self-hosting	OpenAI is easiest

Findings#

1. bce-embedding + the matching reranker takes the lead: alone it ranks 4th but pairing with bce-reranker-base_v1 lifts the combined score to 0.88, just behind bge-m3
2. OpenAI shines on language mixing: when English appears in Chinese context, OpenAI stays steady while open models can dip
3. Dimensions aren’t everything: 3072-dim OpenAI loses to 1024-dim bge-m3

How to plug into Dify#

# Dify → Settings → Model Provider → Add
# Choose Hugging Face or SiliconFlow
# Model: BAAI/bge-m3
# Endpoint: https://api.siliconflow.cn/v1