Qwen3-VL-Embedding-2B

Qwen3-VL-Embedding-2B, published on 7 January 2026, is the small end of Alibaba's multimodal retrieval suite and the model most likely to be the first stage of a working search system. It turns a document into a vector so that a search engine can retrieve by distance rather than by keyword, and the document may be a paragraph, a photograph, a screenshot, a scanned page or a video clip — all mapped into one shared space, so a text query can find a picture and a picture can find a paragraph. It is Qwen3-VL-2B-Instruct adapted for representation: 28 layers, hidden dimension 2,048, a 24-block vision tower fusing into the language model from layers 5, 11 and 17, 2.13 billion parameters, a supported context of 32,768 tokens and more than 30 languages. Output vectors are up to 2,048 numbers long and the length can be cut as far as 64 through Matryoshka training, with quantised embeddings also supported. For a large index that matters more than it sounds: shorter vectors mean a smaller index and faster comparison, and the trade-off can be tuned after the model is already in place. On the vendor's multimodal benchmark MMEB-V2 it scores 73.4 overall, against 77.9 for the 8-billion version. On the text-only MMTEB it scores 63.87 — below Alibaba's own text-only Qwen3-Embedding-4B at 69.45, a model of roughly twice the size but no vision at all. The reading for a practitioner is consistent across this whole family: multimodal indexing is worth having when the archive genuinely contains images, and a straightforward loss when it does not. One further comparison is worth carrying into a design decision. On retrieval from scanned and rendered documents this model is clearly behind the reranker of identical size — 52.9 against 60.8 on ViDoRe v3, 71.0 against 80.9 on JinaVDR. Those figures are not an argument against the embedding model, which does a different job, but they do argue for adding the second stage rather than for buying a larger first one. Adoption in the 30 days to 1 September 2026: 1.38 million Hugging Face downloads and 446 likes — more than the 8-billion embedding model. Weights are Apache 2.0 with no European carve-out.

▸Videos

No videos yet.

▸Reviews