Qwen: Qwen3 Embedding 8B

Qwen3-Embedding-8B is Qwen’s 8B text-embedding model, built from the dense Qwen3 foundation model. It represents text as numerical vectors rather than generating conversational answers. Qwen lists text retrieval, code retrieval, classification, clustering and bitext mining among the supported application areas. A search system can use these vectors to compare a question with relevant passages; a document collection can use them to find semantically related material or organise records into groups.

The exact model supports a 32K-token sequence length and embeddings of up to 4,096 dimensions. It also supports user-defined output dimensions from 32 to 4,096 through Matryoshka Representation Learning. That makes vector length a deliberate design choice: a shorter vector changes the representation’s size, while the full output preserves the model’s maximum dimensionality. Keep the chosen dimension consistent for the documents and queries compared in one index. Qwen3-Embedding-4B is a separate edition with a different maximum dimension, rather than an interchangeable vector source for an existing 8B collection.

The producer describes support for more than 100 languages, including programming languages, and multilingual, cross-language and code retrieval. A practical collection can contain documents in several languages while queries seek the relevant meaning, or source-code passages indexed for a natural-language question. Preserve important identifiers, titles and domain terms in the input. For code, keep the function or module’s surrounding explanation with the excerpt when that context is part of the question. The output remains a vector representation; returning the matched document text is the search application’s next step.

The model is instruction-aware. Qwen recommends a one-sentence instruction tailored to the retrieval task and, for multilingual use, recommends writing that instruction in English. Its direct example formats a query with Instruct: followed by the task description and Query: followed by the question. Retrieval documents do not need the same instruction prefix in that example. A suitable original brief might describe finding technical passages that explain a user’s error, then place the particular error description in the query. The instruction describes what kind of semantic relationship the search should favour.

Qwen’s examples encode queries and documents separately, compare their vectors using cosine similarity, and show a lower-level procedure that pools the final token’s hidden state and normalises the embedding before a dot-product comparison. A useful workflow keeps each passage’s vector attached to its source identifier, compares a prepared query with the stored vectors and returns the corresponding passages in similarity order. The example’s 8,192-token truncation setting is a chosen input-processing setting, distinct from the model’s documented 32K sequence length. Review what is retained when preparing long inputs.

The embedding and reranking models are different components in Qwen’s release. An embedding model represents one text segment and enables an initial similarity search. A separate reranker evaluates the relationship between a query and candidate passages and can refine their order. Qwen describes combining the two modules, but Qwen3-Embedding-8B itself returns embeddings rather than a reranker’s relevance judgement or a written answer. For retrieval followed by an answer, retain the matched source text and pass that evidence to the answering stage. Choose and keep a consistent model, dimension and query instruction for the collection being searched.