Qwen3-Embedding

Qwen3-Embedding, released June 2025, is Alibaba's open-source text embedding series in three sizes: 0.6B (1024 dims, ~1.2GB, CPU-runnable), 4B (2560 dims), and 8B (4096 dims). Trained with a three-stage contrastive pipeline, the series tops Hugging Face embedding benchmarks, supports 100+ languages with particular Chinese-English strength, handles 32K-token contexts, and uses Matryoshka Representation Learning for truncatable dimensions. The 0.6B is a popular Ollama/llama.cpp default for local RAG.

The honest picture: the 8B needs a real GPU (16GB), and while multilingual coverage is broad, low-resource languages lag the headline 100+ number. Alibaba also sells text-embedding-v3/v4 through DashScope, which are newer and API-only — the open Qwen3-Embedding line is the self-hostable tier, not necessarily the absolute best Alibaba offers. For pure English retrieval, BGE and Nomic remain credible alternatives.