Choose a model for semantic search, RAG retrieval, cross-modal matching, and reranking.
Model selection
Choose a model based on your input data type and use case.
- Text or code — Use
text-embedding-v4. Best general-purpose text embedding model, supporting task instructions (instruct), sparse vectors, and other advanced features. - Multimodal content — In Singapore, use
tongyi-embedding-vision-plusfor cross-modal retrieval (text-to-image, image-to-image, text-to-video). For cost-sensitive scenarios, usetongyi-embedding-vision-flash. In China North 2 (Beijing), useqwen3-vl-embedding, which returns fused or independent vectors for text, images, and videos. - Large-scale data — Use
qwen3.7-text-embeddingortext-embedding-v4with Batch API to significantly reduce costs.
How many dimensions?
Large-scale search where storage matters → 256 or 512.
General use → 1024 (default, good balance).
Maximum accuracy on benchmarks → 1536 or 2048.
Text-only data?
Use text-embedding-v4 instead — faster, cheaper, more dimension options.
Multimodal embedding is for cross-modal retrieval (text-to-image, text-to-video).
Accuracy vs speed
Best accuracy → tongyi-embedding-vision-plus (dimensions up to 1152).
Budget or latency-sensitive → tongyi-embedding-vision-flash (up to 768).
Reranking
Improve RAG precision → add qwen3-rerank after your embedding search. Re-scores top-N results with cross-attention for better ranking quality.
Limits: 500 documents per request, 4,000 tokens per item, 120,000 tokens per request.
To rerank images or videos, use qwen3-vl-rerank (China North 2 (Beijing) only). It reranks mixed text, image, and video candidates and is called through the DashScope SDK or HTTP endpoint, not the OpenAI-compatible endpoint. Limits: 100 text documents, 40 images, or 4 videos per request; 8,000 tokens per document; 120,000 tokens per request.
Migrate from closed-source models
Replacing OpenAI, Cohere, or Voyage embeddings? Use these QwenCloud equivalents.
| Task | Closed-source examples | QwenCloud recommendation |
|---|---|---|
| Text embedding | OpenAI text-embedding-3-large, Voyage-3-large, Cohere embed-v4 | text-embedding-v4 |
| Multimodal embedding | Cohere embed-v4, Voyage multimodal | tongyi-embedding-vision-plus |
| Reranking | Cohere Rerank 3.5, Voyage rerank-2.5 | qwen3-rerank (Singapore, China North 2 (Beijing)), qwen3-vl-rerank (China North 2 (Beijing)) |
All models
In the console, the model list, search results, versions, prices, and rate limits follow the current region. Models differ by region, and some are available only in specific regions, so you can only compare models within the same region. Before you choose a model, check the region's tab in Model prices. See Regions and endpoints.
| Model | Regions | Use this when... | Dimensions | Max tokens |
|---|---|---|---|---|
text-embedding-v4 | Singapore, China North 2 (Beijing), China (Hong Kong) | Text search, RAG, clustering | 64, 128, 256, 512, 768, 1024 (default), 1536, 2048 | 8,192 |
text-embedding-v3 | Singapore | Existing v3 index migration | 512, 768, 1024 (default) | 8,192 |
tongyi-embedding-vision-plus | Singapore | Cross-modal search, best accuracy | 1152, 1024, 512, 256, 128, 64 | 1,024 |
tongyi-embedding-vision-flash | Singapore | Cross-modal search, budget | 768, 512, 256, 128, 64 | 1,024 |
qwen3-vl-embedding | China North 2 (Beijing) | Multimodal search with fused or independent vectors | 2560 (default), 2048, 1536, 1024, 768, 512, 256 | 32,000 |
qwen3-rerank | Singapore, China North 2 (Beijing) | Re-rank search results | — | 4,000/item |
qwen3-vl-rerank | China North 2 (Beijing) | Re-rank text, image, and video results | — | 8,000/item |