Qwen3-Embedding-0.6B

Qwen3-Embedding-0.6B, published on 5 June 2025, does not answer questions. It reads a passage and returns 1,024 numbers describing it, so that a search engine can find related material by measuring distance between those numbers. Every document in an index passes through a model like this one, usually once; every query passes through it again. That is why size matters here in a way it does not for a chatbot — and why this, the smallest member of the family, is the one people actually run. It carries 596 million parameters across 28 layers, with a hidden dimension of 1,024, and works over a 32,768-token window. The output vector can be shortened on demand to anything between 32 and 1,024 numbers without retraining, so a project short on memory can trade accuracy for a smaller index. The model also accepts a task instruction prepended to the input; Alibaba measures that at 1 to 5 percent accuracy and advises writing the instruction in English even when the documents are not. The benchmark result worth knowing is the English one. On MTEB English v2 it scores 70.70, ahead of NV-Embed-v2 (7.8 billion parameters, 69.81) and GritLM-7B (67.07). A model that fits in roughly 1.2 GB of memory outranks retrieval models thirteen times its size. On the multilingual leaderboard it scores 64.33, comfortably ahead of the previous small-model standard multilingual-e5-large-instruct (63.22) at the same size. There is a case against it, and the vendor's own table contains it. On the Chinese benchmark C-MTEB it scores 66.33, while ritrieve_zh_v1 — a specialised Chinese model with 0.3 billion parameters, half the size — scores 72.71. For a single-language index, a specialist beats a polyglot, even one from the company that publishes both languages' leading chat models. One detail of how it was trained is unusual enough to state plainly: the first and largest training stage used text pairs generated by the Qwen3 chat model itself, rather than relevance pairs harvested from web forums, which is how such datasets were customarily assembled. Weights are Apache 2.0, with a separate GGUF release for running on a processor without a graphics card. Adoption in the 30 days to 1 September 2026: 6.78 million Hugging Face downloads and 1,174 likes — the twelfth most downloaded model Alibaba publishes.

▸News

Alibaba has shipped three chat generations since June 2025 and not one new text search model — and that is not neglect

9/2/2026 · Qwen (Alibaba Cloud) i Hugging Face

▸Videos

No videos yet.

▸Reviews