Qwen3-Reranker-8B
Alibaba Cloud
The largest reranker in Alibaba's Qwen3 Embedding series: an 8B-parameter, 36-layer cross-encoder built on the Qwen3 foundation model. It scores a query-document pair by reading the model's yes/no token logits under a judging prompt, and is instruction-aware, so a task description tailored to a scenario or language can be supplied per request. It covers over 100 natural and programming languages and is normally paired with a Qwen3-Embedding model as the first-stage retriever. Weights ship on Hugging Face in Safetensors under Apache 2.0 and run through the sentence-transformers CrossEncoder or Transformers; 0.6B and 4B rerankers complete the series.
Strengths
- Apache 2.0 licence on the weights - commercial self-hosting is unrestricted, in contrast to the CC-BY-NC open rerankers it is compared against
- Instruction-aware: a custom task instruction per scenario or language is supported, and the vendor measures a 1-5% retrieval drop when it is omitted
- Coverage of over 100 natural languages plus programming languages, including cross-lingual and code reranking
- Leads the vendor's own comparison table on Chinese (CMTEB-R 77.45), multilingual (MMTEB-R 72.94), long-document (MLDR 70.19) and code (MTEB-Code 81.22) reranking
- Designed to drop in behind the matching Qwen3-Embedding models, giving a fully Apache-2.0 two-stage retrieval stack from one vendor
- Usable through the sentence-transformers CrossEncoder interface or plain Transformers, with 0.6B and 4B variants for lighter deployments
Caveats
- No release date is recorded: the model card states none, and neither the release blog's publication timestamp nor the Hugging Face createdAt stamp is a vendor statement of release
- The 8B does not beat the 4B everywhere in the vendor's own table - it scores 69.02 on MTEB-R against the 4B's 69.76, and 8.05 on FollowIR against the 4B's 14.84, so paying for 2x the parameters can lose English and instruction-following ground
- All reported scores are vendor-run over top-100 candidates retrieved by the vendor's own Qwen3-Embedding-0.6B, not an independent protocol
- Context is advertised only as "32k" / "32K", a rounded label that may be 32000 or 32768, so no context_window integer is recorded here
- Unlike the embedding models, the reranker has no MRL support and no embedding dimension - it emits only relevance scores, so it cannot serve as a first-stage index
- Cross-encoder cost is quadratic in practice: every candidate is re-encoded with the query, so an 8B reranker over long candidate lists is far more expensive to serve than a bi-encoder
- Vendor
- Alibaba Cloud
- License
- Apache-2.0
- Modalities
- text
Capabilities
Resources
Pricing
Open weights under Apache 2.0 - free to download and self-host from Hugging Face in Safetensors with BF16 precision. The vendor publishes no hosted per-token price on the model card or the release blog. As a cross-encoder it scores every query-document pair individually, so serving cost scales with candidate-list length as well as model size.