Qwen3-Reranker-4B

Qwen3-Reranker-4B is a reranking model: it does not search, it re-scores. A search engine returns a rough list of candidate documents, and the reranker reads each query-document pair in full and puts the list in a better order. Alibaba released it on 5 June 2025 alongside the Qwen 3 embedding models, in three sizes meant to be mixed and matched with the embedding side. The maker’s own table has an awkward result in it: on English retrieval (MTEB-R) the 4B model scores 69.76 against 69.02 for the 8B, and on FollowIR it scores 14.84 against 8.05 — nearly double. Bigger is not better here, and the download counts show that practitioners noticed: this size is pulled roughly eleven times more often than the 8B. The 8B keeps the lead on Chinese and multilingual retrieval, so the choice is a question of language mix, not of budget alone. The model is built on the Qwen3-4B-Base checkpoint, handles a 32k-token context and over 100 languages including programming languages, and accepts a written instruction describing what "relevant" means for the task at hand — the maker measures a 1 to 5 percent gain from using one, and advises writing it in English even for other languages, because the training instructions were English. Weights are Apache 2.0, commercial use included; the repository records about 2.53M downloads in the past thirty days.

▸Videos

No videos yet.

▸Reviews