Latest multilingual text embedding from Qwen — default for new projects.
text-embedding-v4 is the latest generation of Qwen's general-purpose multilingual text embedding. It is the recommended choice for new RAG indices, semantic search, clustering, and classification workloads — higher retrieval quality than v3 at the same headline price.
Same OpenAI-compatible `/v1/embeddings` surface as v3, so migration is a model-ID swap plus a one-off reindex. TheRouter routes between bailian-cn and bailian-sg per request based on cost.
| Type | Rate |
|---|---|
| Input | $0.0756 / 1M tokens |
| Release date | 2025-06-05qwenlm.github.io ↗ | verified |
| Parameter count | 4Bhuggingface.co ↗ | verified |
| Transformer layers | 36huggingface.co ↗ | verified |
| Maximum embedding dimensions | 2560 (MRL: 32–2560)huggingface.co ↗ | verified |
| Context length (TheRouter) | 8,192 tokenstherouter.ai ↗ | verified |
| Upstream context length (native) | 32,768 tokenshuggingface.co ↗ | verified |
| Supported languages | 100+ (natural languages + programming languages)github.com ↗ | verified |
| MRL (Matryoshka) support | Yes — output dimensions configurable from 32 to 2560huggingface.co ↗ | verified |
| Instruction-aware prompting | Yes — prepend task description to query for 1–5% retrieval gaingithub.com ↗ | verified |
| License | Apache 2.0github.com ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MTEB Multilingual (Qwen3-Embedding-8B, for context) 8B variant; ranked #1 on MTEB multilingual leaderboard as of June 5, 2025. The 4B variant (text-embedding-v4) scores lower but remains competitive in its size class. | 70.58score | qwenlm.github.io ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "qwen/text-embedding-v4",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Generate embeddings via the OpenAI-compatible /v1/embeddings endpoint. Supports configurable output dimensions (32–2560, Matryoshka MRL) and instruction-aware query prompting for 1–5% retrieval gain. TheRouter auto-routes between Alibaba Bailian bailian-cn and bailian-sg.
curl https://api.therouter.ai/v1/embeddings \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/text-embedding-v4",
"input": "The quick brown fox jumps over the lazy dog",
"encoding_format": "float",
"dimensions": 2560
}'
# MRL: reduce dimensions for lower storage cost (range 32-2560)
# "dimensions": 512
# Instruction-aware query (prepend task prefix to query, NOT to documents):
# "input": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: What is the capital of France?"Alibaba Qwen released the Qwen3-Embedding-0.6B/4B/8B and Qwen3-Reranker-0.6B/4B/8B series on Hugging Face and ModelScope under Apache 2.0. The 8B embedding model claimed the #1 spot on the MTEB multilingual leaderboard with a score of 70.58. The release includes a technical report (arXiv 2506.05176) and open-source training code on GitHub.
Yes. qwen/text-embedding-v4 is the API product name Alibaba Bailian uses for Qwen3-Embedding-4B. The TheRouter model ID qwen/text-embedding-v4 routes to this model. The 8B variant is not separately available through this ID.
No. v4 and v3 vectors live in different embedding spaces and are not comparable. A full reindex of your vector database is required. If reindexing is not immediately feasible, keep calling qwen/text-embedding-v3 for existing indices until you are ready to rebuild.
Pass the dimensions parameter in your embeddings request (e.g. dimensions: 512). The valid range is 32–2560. Smaller dimensions reduce storage and ANN query cost with a moderate quality trade-off; benchmark on your specific dataset before choosing a reduced dimension for production.
No. The recommended pattern is to prefix only your query with the task instruction (e.g. Instruct: Given a web search query, retrieve relevant passages...\nQuery: <your query>). Passage/document inputs should be sent without any instruction prefix. Prepending instructions to documents typically hurts retrieval quality.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Parameter count | huggingface.co ↗ | 2026-06-09 | verified |
| Transformer layers | huggingface.co ↗ | 2026-06-09 | verified |
| Maximum embedding dimensions | huggingface.co ↗ | 2026-06-09 | verified |
| Context length (TheRouter) | therouter.ai ↗ | 2026-06-09 | verified |
| Upstream context length (native) | huggingface.co ↗ | 2026-06-09 | verified |
| Supported languages | github.com ↗ | 2026-06-09 | verified |
| MRL (Matryoshka) support | huggingface.co ↗ | 2026-06-09 | verified |
| Instruction-aware prompting | github.com ↗ | 2026-06-09 | verified |
| License | github.com ↗ | 2026-06-09 | verified |
| Training cutoff | — | — | unknown |
| MTEB Multilingual (Qwen3-Embedding-8B, for context) | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Qwen3 Embedding series open-sourced under Apache 2.0 | qwenlm.github.io ↗ | 2026-06-09 | verified |
| Is text-embedding-v4 the same as Qwen3-Embedding-4B? | bailian.console.aliyun.com ↗ | 2026-06-09 | to verify |
| Can I migrate from text-embedding-v3 without reindexing? | therouter.ai ↗ | 2026-06-09 | to verify |
| How do I use MRL to reduce vector dimensions? | huggingface.co ↗ | 2026-06-09 | to verify |
| Should I add an instruction prefix to my documents as well as my queries? | github.com ↗ | 2026-06-09 | to verify |