Skip to main content
Back

Alibaba's Qwen3-Embedding-8B Tops MTEB and Beats Google's Proprietary API

Alibaba's Qwen3-Embedding-8B tops the MTEB multilingual leaderboard with a 70.58 score, ships Apache 2.0, and covers 100+ languages at 32K context.

  • Qwen3-Embedding-8B tops MTEB multilingual with a 70.58 score, beating Gemini embeddings.
  • Apache 2.0 licensed, 8B params, 32K context, up to 4096-dim vectors with MRL truncation.
  • Supports 100+ languages and programming languages, plus instruction-aware queries for task steering.
  • Built on Qwen3-8B-Base via LoRA, dual-encoder architecture with EOS-token pooling.
  • Trained in three stages: synthetic contrastive pretraining, supervised fine-tuning, then model merging.
  • Sibling 0.6B and 4B embedders plus matching rerankers available on GitHub.

Alibaba's Qwen team dropped a serious contender into the open embedding space with Qwen3-Embedding-8B, an 8-billion parameter text embedding model built on the Qwen3 foundation series. It launched at the top of the MTEB multilingual leaderboard and has already crossed 2.4 million downloads on Hugging Face, signaling that a lot of retrieval pipelines are quietly getting rewired around it.

The series targets text embedding and ranking tasks, building on the dense Qwen3 base models and shipping in three sizes: 0.6B, 4B, and 8B. The 8B variant is the flagship, and it comes paired with a matching reranker family so you can swap in a stronger second stage without leaving the ecosystem.

Leaderboard numbers that actually matter

As of June 2025, the 8B model sits at #1 on the MTEB Multilingual leaderboard with a mean task score of 70.58 and a mean type score of 61.69, released under Apache 2.0. It beats gte-Qwen2-7B-Instruct (62.51) and edges out Google's gemini-embedding-exp-03-07 (68.37) on the same suite.

English-only results follow the same shape. On MTEB (Eng v2) it hits 75.22 mean task and 68.71 mean type, ahead of NV-Embed-v2 (69.81), GritLM-7B (67.07), and stella_en_1.5B_v5 (69.43). On C-MTEB (Chinese) it reaches 73.84, ahead of gte-Qwen2-7B-instruct at 71.62. Independent benchmarks back this up: on Polish MTEB, the 8B model took the overall top spot, with its advantage over other models (including Polish-specific ones) driven mainly by classification and clustering.

What the spec sheet gives you

  • Parameters: 8B, 36 layers, BF16 weights
  • Context length: 32K tokens
  • Embedding dimension: up to 4096, user-configurable from 32 to 4096 via Matryoshka Representation Learning (MRL)
  • Languages