Qwen3-VL-Reranker-8B

Qwen3-VL-Reranker-8B, published on 7 January 2026, occupies the second stage of a search pipeline. An embedding model first pulls perhaps a hundred candidates out of a large index quickly and roughly; the reranker then reads the query together with each candidate as one input and issues a relevance score. Because it sees both sides at once rather than comparing pre-computed vectors, it judges far better — and costs far more, since the work has to be repeated for every candidate of every query. The input can be text, an image, a screenshot, a slide or a video frame, in any mixture, on either side of the pair. That is the point of the model: a query written in words can be scored against a page that exists only as a scan. It is built on Qwen3-VL-8B-Instruct: 36 layers, hidden dimension 4,096, a 27-block vision tower, 8.77 billion parameters and a supported context of 32,768 tokens. The parameter count is about 622 million higher than the embedding model of the same family, and the difference is exactly the size of the vocabulary output layer — 151,936 tokens by 4,096 dimensions. That layer is not decoration here: the reranker produces its score as a logit over an output token, so it needs the head that the embedding model was able to discard. On every column of the comparison Alibaba published it is the best of the four: 79.2 average on the MMEB-v2 retrieval split, 86.3 on visual document retrieval, 74.9 on MMTEB retrieval, 83.6 on JinaVDR and 66.7 on ViDoRe v3 — the last figure roughly six points above the 2-billion reranker and nearly fourteen above the 2-billion embedding model. And almost nobody runs it. In the 30 days to 1 September 2026 it was downloaded 43,201 times from Hugging Face, against 2.07 million for Qwen3-VL-Reranker-2B: a ratio of about one to forty-eight. The explanation is structural rather than a matter of taste. A reranker is invoked once per candidate, so a hundred-document shortlist means a hundred forward passes before the user sees anything, and at eight billion parameters that is a budget most search systems will not spend. This model is the right choice where correctness outranks latency — legal discovery, medical archives, technical documentation — and the wrong one behind a search box. Weights are Apache 2.0 with no European carve-out. The suite is described in a technical report published in January 2026 (arXiv 2601.04720).

▸Videos

No videos yet.

▸Reviews