Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking

Enterprise search increasingly covers more than text: product photos, scanned contracts, slide decks, UI screenshots, training videos. The Qwen team at Alibaba released two model families for this, both built on the Qwen3-VL vision-language model and both in 2B and 8B sizes.

The embedding model maps any input (text, an image, a document page, a video, or a mix) to a single vector in a shared space, so a text query can find a matching image or video clip. It handles inputs up to 32,000 tokens, supports more than 30 languages, and is trained so its vectors can be truncated to fewer dimensions (Matryoshka learning) and stored at low precision. The reranker reads a query and a candidate together and scores relevance as the probability of answering "yes" to whether the document meets the query.

The training story is the interesting part. The embedder is trained in three stages: contrastive pre-training on 300 million synthetic pairs, multi-task training on curated data, then distillation from the reranker, so the cheap model learns the expensive model's judgments. A final merge recovers ground lost on non-retrieval tasks.

The results: the 8B embedder ranked first on the MMEB-V2 multimodal benchmark at 77.8. On visual document retrieval, a single vector per page came within two points of the best multi-vector models, and reranking the top 100 candidates beat them all.

Treat this as a reference design for production search, with the trade-offs measured.

Retrieve wide, rerank close. An embedding model is fast and approximate; a reranker is slow and precise. On visual documents, the 8B embedder alone scored 75.8, and reranking its shortlist scored 80.3. Running a cross-encoder over 100 candidates per query costs real GPU time, so the reranking depth is a latency and budget decision you should make with numbers.

Dimensions and precision set the storage bill. Cutting dimensions from 1,024 to 512 cost about 1.4% in retrieval quality and halved storage, roughly doubling search speed. Storing vectors as int8 was close to lossless; binary vectors hurt a lot, and more so at lower dimensions. For a corpus of tens of millions of items, these settings change hardware costs more than the choice of model.

One model for everything is a compromise. The multimodal 8B embedder scored 67.9 on a text-only benchmark, below the text-only Qwen3-Embedding-8B at 70.6. If the customer's corpus is almost all text, a text model may still be the better pick. If it mixes text with pages, images and video, one shared space removes a lot of plumbing.