Machine-translated from the Korean original. Read the original

Qwen3-Embedding 0.6B Embedding Performance

Contents

What We Know

The Qwen3-Embedding series is a family of text embedding and reranking models released by Alibaba's Qwen team in June 2025. It comes in three sizes — 0.6B / 4B / 8B — and this note summarizes the performance of the smallest, the 0.6B version.

Basic Specifications

Item Value
Parameters ~600 million (0.6B), 28 transformer layers
Context length 32K tokens
Embedding dimensions Up to 1024 (user-configurable between 32 and 1024, via MRL)
Supported languages 100+ (natural languages + programming languages)
Instruction-aware Yes (accepts task-specific instructions as input)

Benchmark Performance

MTEB Multilingual (MMTEB)

Model Size Average Score
Qwen3-Embedding-0.6B 0.6B 64.33
multilingual-e5-large-instruct 0.6B 63.22
Cohere embed-multilingual-v3.0 - 61.12
Gemini Embedding - 68.37
Qwen3-Embedding-4B 4B 69.45
Qwen3-Embedding-8B (leaderboard #1, 2025-06) 8B 70.58

The 0.6B model edges out the leading open-source model of comparable size (multilingual-e5-large-instruct) and also comes in above Cohere's commercial multilingual embedding.

Comparison with OpenAI (text-embedding-3)

Item Qwen3-Embedding-0.6B text-embedding-3-small text-embedding-3-large
MTEB (English) 70.70 (v2) 62.3 64.6
Multilingual bench (MTEB/MIRACL) 64.33 (MMTEB) 44.0 (MIRACL) 54.9 (MIRACL)
Max context 32K 8K 8K
Embedding dimensions 32–1024 (MRL) 512 / 1536 256 / 1024 / 3072
Price Free (local) $0.02 / 1M tokens $0.13 / 1M tokens
Availability Open weights (Apache 2.0) Closed API Closed API

Caveat: OpenAI's official scores are based on older versions of MTEB/MIRACL, while the Qwen3 numbers are based on MMTEB v2, so the axes aren't perfectly aligned. That said, the finding that a 0.6B open model beats text-embedding-3-large on both English retrieval and multilingual tasks is reproduced across several independent leaderboards.

Graph
Graph