Qwen3-Embedding-8B

Qwen3-Embedding-8B, published on 5 June 2025, is the accuracy ceiling of Alibaba's text retrieval line. It turns a passage into a vector of up to 4,096 numbers; a search engine then finds related documents by measuring distance between vectors. On the day of release it took first place on the MTEB multilingual leaderboard with a mean of 70.58, and it holds the highest score in every one of the vendor's three benchmark tables: 70.58 multilingual, 75.22 on MTEB English v2, 73.84 on the Chinese C-MTEB. It is built on Qwen3-8B-Base and keeps its shape: 36 layers, hidden dimension 4,096, a 32,768-token window. The weights file holds 7.57 billion parameters rather than the 8.19 billion of the base chat model, and the missing 621 million are not a smaller model — they are the output layer that converts hidden states back into vocabulary tokens, which a model producing only vectors never uses. Alibaba left it out of the file. Output vectors shorten on demand to as few as 32 numbers, and the model accepts a task instruction prepended to the input. The reason to think twice is cost rather than quality. Indexing runs this model over every document in a collection, so choosing the 8B over the 0.6B multiplies the indexing bill by roughly thirteen for a gain of 6.25 points of multilingual MTEB — and the download figures show which trade-off practitioners make: 2.35 million downloads for this model against 6.78 million for the small one. One comparison inside Alibaba's own catalogue is worth stating. Qwen3-VL-Embedding-8B, published seven months later and able to index photographs and video, scores 67.88 on multilingual text — 2.7 points below this older text-only model of identical size. For an index containing nothing but text, the newer model is not an upgrade. Weights are Apache 2.0, with a separate GGUF build. Adoption in the 30 days to 1 September 2026: 2.35 million Hugging Face downloads and 788 likes.

▸Videos

No videos yet.

▸Reviews