Qwen3-Embedding-8B

Qwen3-Embedding-8B is a specialized text embedding model built upon the dense foundational architecture of the Qwen3 series. Designed specifically for embedding and ranking tasks, it produces high-quality vector representations of text with flexible dimensionality ranging from 32 to 4096 dimensions, allowing developers to tune output size based on their storage and accuracy requirements. The model inherits the multilingual strengths of its base, supporting over 100 languages including various programming languages, and handles context windows up to 32K tokens for long-document understanding. Its primary strengths lie in text retrieval, code retrieval, text classification, text clustering, and bitext mining—making it a versatile backbone for semantic search pipelines and information retrieval systems.

The model achieved state-of-the-art performance by ranking No.1 on the MTEB multilingual leaderboard as of June 2025 with a score of 70.58, demonstrating its exceptional capability in multilingual retrieval scenarios. Qwen3-Embedding-8B supports instruction-aware embedding generation, enabling users to guide the model's output for specific tasks, languages, or specialized domains by incorporating custom instructions. This combination of high benchmark performance, flexible dimensionality, multilingual support, and instruction-awareness makes it particularly well-suited for building cross-lingual search systems, enterprise knowledge bases, and AI applications requiring semantic understanding across diverse text corpora and programming code.

Regolo AIqwen3-embedding-8bqwen

Open provider page Provider docs

Quick Info

Provider: Regolo AI
Model key: qwen3-embedding-8b
Release date: Feb 1, 2026
Last updated: Feb 1, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $0.10

Limits

Output tokens: 8,192 tokens
Context window: 32,768 tokens

Latest news about Qwen3-Embedding-8B

Regolo AI

CoverageBenchmark

The MTEB Leaderboard entry for Qwen/Qwen3-Embedding-8B documents concrete, version-explicit technical specifications for the exact model variant: 7.6B total parameters with 6.9B active, a 4,096-dimensional embedding space, a 32,768-token maximum input length, an Apache-2.0 license, and a release date of 2025-06-05. The page also lists the training data sources used (including CMedQAv2-reranking, CodeSearchNet, DuRetrieval, FEVER, HotpotQA, MIRACLRetrieval, MMarcoReranking, MSMARCO, MrTidyRetrieval, NQ, and T2Retrieval) and notes broad multilingual coverage across 100+ languages, confirming the model's positioning as a multilingual text embedding model rather than a single-language retriever. The leaderboard captures per-benchmark ranking snapshots for the exact Qwen3-Embedding-8B variant, including top placements in multilingual retrieval and code retrieval categories (e.g., MTEB (Multilingual), BEIR, Code Information Retrieval, BRIGHT-Pro), along with task-specific mean scores on BEIR, MTEB (Multilingual), RTEB Healthcare, and Long-Horizon Memory. These benchmarks are explicitly tied to the Qwen3-Embedding-8B model card and reference, distinguishing this exact 8B variant from sibling 0.6B and 4B models in the Qwen3 embedding series described elsewhere on the same page.