Qwen3-Embedding-8B
Alibaba Cloud
The largest text embedding model in Alibaba's Qwen3 Embedding series, an 8B-parameter, 36-layer dual-encoder built on Qwen3-8B-Base and fine-tuned with LoRA for embedding and ranking work. It emits vectors of up to 4096 dimensions with user-defined output dimensions from 32 to 4096 (Matryoshka), is instruction-aware so a task description can be prepended to queries, and covers over 100 natural and programming languages for multilingual, cross-lingual and code retrieval. Weights ship on Hugging Face in Safetensors under Apache 2.0 and run through sentence-transformers, Transformers, vLLM or Text Embeddings Inference; 0.6B and 4B siblings cover lower-resource deployments.
Strengths
- Apache 2.0 licence on the weights - unrestricted commercial self-hosting, unlike the CC-BY-NC open embedding models it competes with
- User-defined output dimensions anywhere from 32 to 4096 (MRL support), letting index size be traded against quality without retraining
- Instruction-aware: a per-task, per-language instruction can be prepended to queries, which the vendor measures as a 1-5% gain on most downstream tasks
- Coverage of over 100 natural languages plus programming languages, supporting multilingual, cross-lingual and code retrieval
- Highest scores in the vendor's own tables across MTEB multilingual (70.58), MTEB English v2 (75.22) and C-MTEB (73.84)
- Runs out of the box in sentence-transformers, Transformers, vLLM and Text Embeddings Inference, with matching 0.6B/4B embedders and 0.6B/4B/8B rerankers for a full retrieval stack
Caveats
- No release date is recorded: the vendor states none on the model card, and neither the blog's publication timestamp nor the Hugging Face createdAt stamp is a vendor statement of release
- Bigger is not uniformly better in the vendor's own table: on MMTEB instruction retrieval the 8B scores 10.06, below the 4B's 11.56
- Weak on summarisation in MTEB English v2 (34.83) - behind stella_en_1.5B_v5 (36.91) and gemini-embedding-exp-03-07 (38.28), both far smaller or cheaper to run
- The vendor's "No.1 on the MTEB multilingual leaderboard" claim is timestamped "as of June 5, 2025" and the comparison scores were pulled from the leaderboard on 24 May 2025, so the ranking is over a year stale
- Context is advertised only as "32k" / "32K", a rounded label that may be 32000 or 32768, so no context_window integer is recorded here
- At 8B parameters it is the heaviest option in the series; the vendor positions the 0.6B and 4B variants for deployments "that prioritize efficiency"
- Vendor
- Alibaba Cloud
- License
- Apache-2.0
- Modalities
- text
| Benchmark | Score | Max |
|---|---|---|
| MTEB (Multilingual) Mean (Task) | 70.6 | 100.0 |
| MTEB (Eng v2) Mean (Task) | 75.2 | 100.0 |
| C-MTEB (MTEB Chinese) Mean (Task) | 73.8 | 100.0 |
Capabilities
Resources
Pricing
Open weights under Apache 2.0 - free to download and self-host from Hugging Face (BF16 Safetensors, sharded). The vendor publishes no hosted per-token price on the model card or the release blog; the Hugging Face page lists third-party inference providers instead. Self-hosting an 8B embedding model requires GPU capacity that the 0.6B and 4B variants do not.