TheRouter.aiTheRouter.ai
Back to Models

Embedding 3

zhipuzhipu/embedding-3

Zhipu AI Embedding 3 β€” latest text embedding model. Default 2048-dim, supports custom dimensions (256, 512, 1024, 2048).

Embedding-3 β€” Official Zhipu AI Documentation β†—Zhipu AI Open Platform (bigmodel.cn) β†—

Zhipu AI Embedding-3 is the third-generation text embedding model from Beijing-based Zhipu AI (ζ™Ίθ°± AI), the research lab behind the GLM series of language models. Released as part of the bigmodel.cn open platform, Embedding-3 is a significant step forward from the fixed-dimension Embedding-2, introducing adjustable output dimensions (256 / 512 / 1024 / 2048) and an 8K token input window β€” up from Embedding-2's 512-token limit. The model is designed for Chinese-English bilingual semantic understanding, with particular strength in Chinese-dominant retrieval corpora.

For TheRouter operators, Embedding-3 is the recommended embedding model for Chinese-first RAG, search, and recommendation pipelines. At Β₯0.5/MTok (~$0.069/MTok via TheRouter), it is one of the most cost-efficient embedding APIs in the Chinese-language AI stack. Its matryoshka-style dimension selection allows operators to trade off storage cost and vector search latency β€” for example, using 256-dim vectors for large-scale approximate search and 2048-dim for high-fidelity retrieval β€” without re-embedding the entire corpus. The model is accessible through the OpenAI-compatible /v1/embeddings endpoint at api.therouter.ai using the zhipu/embedding-3 model identifier.

Best for
  • β€’ Chinese-English bilingual RAG pipelines: semantic search over mixed-language knowledge bases where Chinese documents dominate
  • β€’ Cost-sensitive large-scale embeddings: Β₯0.5/MTok pricing makes it the default choice for high-volume Chinese-language corpora (>100M tokens/day)
  • β€’ Dimension-flexible deployments: matryoshka dimension control (256–2048) lets you compress vectors for ANN indexes without full re-embedding
  • β€’ Recommendation systems and content deduplication over large Chinese-language text corpora (news, e-commerce, user reviews)
Reach for something else if
  • β€’ Purely English-language retrieval at scale: OpenAI text-embedding-3-large (3072-dim, MTEB SOTA on English tasks) or Cohere Embed V4 (multimodal, multilingual) are better fits
  • β€’ Multimodal (image + text) embedding tasks: Cohere Embed V4 supports interleaved image/text inputs; Embedding-3 is text-only
  • β€’ Cross-lingual semantic tasks beyond Chinese and English: models with broader multilingual MTEB coverage (e.g., Cohere Embed V4, multilingual-e5-large) are more reliable
Context Length
3K
Max Output
--
Input Priceper 1M tokens
$0.081/ 1M tokens

Modalities

text→embedding

Pricing Breakdown

TypeRate
Input$0.081 / 1M tokens

Supported Parameters

inputdimensionsencoding_formatuser

Specifications

Context window8,192 tokens (8K)docs.bigmodel.cn β†—verified
Output dimensions256, 512, 1024, or 2048 (default: 2048)docs.bigmodel.cn β†—verified
Batch size limitUp to 64 strings per requestdocs.bigmodel.cn β†—verified
List price (Zhipu direct)Β₯0.5 per million tokensdocs.bigmodel.cn β†—verified
Matryoshka dimension reductionSupported β€” select 256/512/1024/2048 via the dimensions parameterdocs.bigmodel.cn β†—verified
ModalityText β†’ embedding (text-only input; no image or audio support)docs.bigmodel.cn β†—verified
Generation (predecessor)Zhipu Embedding-2 (1024-dim fixed, 512-token context)docs.bigmodel.cn β†—verified

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "zhipu/embedding-3",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Embeddings (OpenAI-compatible)

Access Embedding-3 through the standard OpenAI /v1/embeddings endpoint at api.therouter.ai. Pass the dimensions parameter (256, 512, 1024, or 2048) to select your output size. Batch up to 64 strings per request for maximum throughput.

cURL
curl https://api.therouter.ai/v1/embeddings \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zhipu/embedding-3",
    "input": ["ζ™Ίθ°± AI ζ˜―δΈ­ε›½ι’†ε…ˆηš„ AI η ”η©ΆζœΊζž„", "Zhipu AI is a leading Chinese AI research lab"],
    "dimensions": 1024
  }'

More from zhipu

zhipu/embedding-2zhipu/glm-4-flashzhipu/glm-4-air

Similar models

Cross-provider sibling models
openai/text-embedding-3-small
OpenAI's fast embedding model β€” stronger English MTEB; 8K context; 1536-dim default
openai/text-embedding-3-large
OpenAI's highest-quality embedding model β€” 3072-dim, best English MTEB
cohere/embed-v4
Cohere's multimodal multilingual embedding model β€” supports image+text, binary quantization

Frequently asked

Which dimension should I use?

2048 (default) delivers the highest semantic fidelity and is recommended for precision-critical retrieval such as legal or medical knowledge bases. 1024 is a practical balance for most production RAG stacks. 512 suits large-scale approximate nearest-neighbor (ANN) search where index storage cost matters. 256 is suitable for real-time similarity scoring where latency is the primary constraint. All four dimensions share the same token cost per request.

Can I use Embedding-3 with LangChain, LlamaIndex, or Spring AI?

Yes. Because TheRouter exposes Embedding-3 through the standard OpenAI embeddings API at api.therouter.ai, any framework that supports the OpenAI embeddings client works out of the box. For LangChain and LlamaIndex, set the base URL to https://api.therouter.ai/v1 and the model to zhipu/embedding-3. For Spring AI, set spring.ai.openai.embedding.options.model=zhipu/embedding-3 and spring.ai.openai.embedding.options.dimensions=1024 (or your preferred size).

How does Embedding-3 compare to OpenAI text-embedding-3-small?

Embedding-3 is significantly more cost-efficient (~0.069/MTok vs ~$0.020/MTok) but targets Chinese-first workloads. OpenAI text-embedding-3-small is stronger on English-only MTEB benchmarks and is the default for English-dominant RAG stacks at TheRouter. Embedding-3's 8K context window (vs 8,191 tokens for text-embedding-3-small) and matryoshka dimension control are comparable. Choose Embedding-3 when Chinese recall quality is the primary concern; choose text-embedding-3-small for English or mixed Western-language corpora.

Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Context windowdocs.bigmodel.cn β†—2026-06-09verified
Output dimensionsdocs.bigmodel.cn β†—2026-06-09verified
Batch size limitdocs.bigmodel.cn β†—2026-06-09verified
List price (Zhipu direct)docs.bigmodel.cn β†—2026-06-09verified
Matryoshka dimension reductiondocs.bigmodel.cn β†—2026-06-09verified
Modalitydocs.bigmodel.cn β†—2026-06-09verified
Generation (predecessor)docs.bigmodel.cn β†—2026-06-09verified
Help & contact