Qwen3-Embedding

Alibaba's Qwen team released Qwen3-Embedding in June 2025 as a series of embedding and reranking models in 0.6B, 4B, and 8B sizes, all built on the Qwen3 foundation models with 32K-token context windows. The 8B embedder debuted at number one on the MTEB multilingual leaderboard with a score of 70.58, and the series covers more than 100 natural languages plus programming languages, with instruction-aware encoding that lets task prompts steer retrieval behavior. Embedding dimensions run 1024, 2560, and 4096 by model size, and Matryoshka Representation Learning allows user-defined output dimensions from 32 to 4096. The matching Qwen3-Reranker models score query-document pairs for second-stage ranking. Weights are Apache-2.0 on Hugging Face, and inference works through Transformers 4.51 or newer, vLLM 0.8.5 or newer, or Sentence Transformers 2.7 or newer; the 0.6B models are practical on CPU while the larger variants want a GPU. Retrieval-augmented generation, multilingual search, and code search are the primary deployment targets.