A 23M-Parameter Model Still Wins: 338 Embedding Models Ranked by Real Downloads

本文由「A 23M-Parameter Model Still Wins: 338 Embedding Models Ranked by Real Downloads」的公开内容整理，完整内容与后续更新请点击文末链接阅读原文。

New StoryA 23M-Parameter Model Still Wins: 338 Embedding Models Ranked by Real DownloadsbyWritings, Papers and Blogs on Text ModelsbyWritings, Papers and Blogs on Text Models|@textmodelsWe publish the best academic papers on rule-based techniques, LLMs, & the generation of text that resembles human text.November 12th, 2026TLDR Your browser does not support the audio element.Speed1xVoiceDr. One Ms. Hacker byWritings, Papers and Blogs on Text Models@textmodelsbyWritings, Papers and Blogs on Text Models|@textmodelsWe publish the best academic papers on rule-based techniques, LLMs, & the generation of text that resembles human text.Story's CredibilitybyWritings, Papers and Blogs on Text Models|@textmodelsWe publish the best academic papers on rule-based techniques, LLMs, & the generation of text that resembles human text.Story's Credibilityall-MiniLM-L6-v2, a 23M-parameter sentence-transformers model, was downloaded 177M times on Hugging Face in the last 30 days. That's 38% of all downloads across the 213 open-weight embedding models tracked by the TextModels directory, and more than the next four models combined. Below are all 338 embedding models in the directory: open weights ranked by downloads, then hosted APIs, Ollama builds and the historical models that started it all. Most "best embedding model" lists rank by benchmark score, like PremAI's ranking by MTEB score, cost and self-hosting. Benchmarks tell you what's best on paper. Downloads tell you what developers actually ship, and the two lists look very different. The sentence-transformers family's lead has a history. As HackerNoon's overview of embeddings for RAG puts it, "Simply said, the BERT model isn't the best for similarity search." SBERT fixed that by embedding each sentence separately so vectors can be computed once and compared fast, and its descendants still dominate the download charts. What the download data says Small and old beats big and new. The top two models by downloads, all-MiniLM-L6-v2 (23M params) and BAAI's bge-small-en-v1.5 (33M), are tiny. For most production search, "fast and good enough" wins. The market is concentrated. The top 10 open-weight embedders account for 74% of the 470M monthly downloads in this list. Qwen is the newcomer to watch. Qwen3-Embedding-0.6B is already #9 with 9M monthly downloads, and its 4B and 8B siblings both clear 2.4M. Multilingual is mainstream. Five of the top 12 are multilingual: bge-m3, two paraphrase-multilingual models and two multilingual-e5 models. Vector databases ship their own builds. Qdrant's ONNX conversions and sparse models appear nine times, which shows how many teams get embeddings through their database instead of picking a model directly. How this list was built Everything comes from the TextModels directory's Embedding filter (338 models) as of November 11, 2026. Hugging Face entries are ranked by 30-day downloads pulled from the Hugging Face API. "Verified" means the TextModels directory fetched and confirmed that model's metrics per model. Hosted API models have no public download counts, so they're grouped by provider, alphabetically. Some hosted entries, like OpenAI's 001-series search models, are long retired. They stay in the list because people still search for them when migrating old pipelines. Part 1: 213 open-weight embedding models on Hugging Face, ranked by 30-day downloads all-MiniLM-L6-v2 by sentence-transformers: 177M downloads/30d, 23M params, Apache 2.0 (Verified) bge-small-en-v1.5 by BAAI: 47.1M downloads/30d, 33M params, MIT (Verified) bge-m3 by BAAI: 29.3M downloads/30d, 568M params, MIT (Verified) paraphrase-multilingual-MiniLM-L12-v2 by sentence-transformers: 26.9M downloads/30d, 118M params, Apache 2.0 (Verified) all-mpnet-base-v2 by sentence-transformers: 20.5M downloads/30d, 109M params, Apache 2.0 (Verified) nomic-embed-text-v1.5 by nomic-ai: 11.9M downloads/30d, 137M params, Apache 2.0 (Verified) multilingual-e5-small by intfloat: 9.4M downloads/30d, 118M params paraphras