Qwen3-Embedding

weights

open(three ungated Apache-2.0 checkpoints, 0.6B/4B/8B)

data

closed(multi-stage weakly-supervised and synthetic mixture described in the paper, not released)

code

partial(evaluation and inference examples in QwenLM/Qwen3-Embedding

license

Apache-2.0(OSI)

Apache-2.0 across every distributed size, which is unusual for a Qwen release and means multi_sku_rule has nothing restrictive to resolve to. What holds it at 3 rather than higher is the other half of the ladder: the training corpus is described in the paper but not published, and the repository carries evaluation and usage code rather than the pipeline that produced the checkpoints.

• https://huggingface.co/Qwen/Qwen3-Embedding-8B/raw/main/README.md recorded 2026-09-11

  license: apache-2.0 in the front matter; base_model Qwen/Qwen3-8B-Base; the three embedding sizes and three reranker sizes listed as downloadable; no training data or pretraining code named.

• https://raw.githubusercontent.com/QwenLM/Qwen3-Embedding/main/README.md recorded 2026-09-11

  The product repository points at the Hugging Face and ModelScope collections, the blog, the arXiv paper and the hosted API, and offers usage and evaluation material rather than a pretraining pipeline.

• https://huggingface.co/api/models/Qwen/Qwen3-Embedding-8B recorded 2026-09-11

  gated: false; private: false; cardData license apache-2.0.

12,549,679 downloads in the trailing 30 days across the three shipped embedding checkpoints - 7,806,497 for the 0.6B, 2,489,577 for the 4B and 2,253,605 for the 8B. Excluded from the sum: the vendor's own GGUF conversions of the same three checkpoints (another 173k) and the separate Qwen3-VL-Embedding line, which is a different product. Level 5 here is measured, not inferred.

• https://huggingface.co/api/models/Qwen/Qwen3-Embedding-0.6B recorded 2026-09-11

  downloads: 7806497; gated: false; private: false.

• https://huggingface.co/api/models/Qwen/Qwen3-Embedding-4B recorded 2026-09-11

  downloads: 2489577; gated: false; private: false.

• https://huggingface.co/api/models/Qwen/Qwen3-Embedding-8B recorded 2026-09-11

  downloads: 2253605; gated: false; private: false.

Rung 4, the competitive frontier. Qwen3-Embedding-8B held first place on the multilingual MTEB leaderboard at release and still posts 70.58, with 100+ languages, a 32k context and Matryoshka dimensions - everything the rung asks for. It is one below Harrier-OSS because the anchor's 74.3 has since displaced it from the top of the same table, not because anything about the Qwen line has weakened.

• https://huggingface.co/Qwen/Qwen3-Embedding-8B/raw/main/README.md recorded 2026-09-11

  The card states the 8B model ranks No.1 in the MTEB multilingual leaderboard as of June 5, 2025 with a score of 70.58, over 100 languages, 32k context and embedding dimensions up to 4096.

• https://huggingface.co/nvidia/llama-embed-nemotron-8b/raw/main/README.md recorded 2026-09-11

  A third-party MMTEB table dated 21 October 2025 places Qwen3-Embedding-8B third by Borda rank with the highest mean task score in the table, 70.58 - corroborating the vendor figure from outside.