PyTorch Model Zoo Stats

As of April 2026, the most downloaded PyTorch model is sentence-transformers/all-MiniLM-L6-v2 with over 194 million downloads. This sentence embedding model is popular for semantic search, clustering, and retrieval-augmented generation (RAG) pipelines. Google's BERT-base-uncased ranks second with 66M+ downloads.

Sentence embedding models dominate the top downloads, with sentence-transformers appearing 4 times in the top 20. Text generation models (Qwen3, Llama 3.1, GPT-2) are the second most popular category. CLIP models for zero-shot image classification hold strong in the top 10. The trend shows a shift toward specialized embedding and multimodal models.

Consider: 1) Task match — use the pipeline_tag to find models for your task, 2) Model size vs. hardware — MiniLM (22M params) runs on CPU while Llama-3.1-8B needs a GPU, 3) License — Apache 2.0 and MIT are commercially permissive, 4) Community validation — higher likes and downloads indicate battle-tested models, 5) Recency — newer architectures like ModernBERT often outperform older ones.

Apache 2.0 is the most common license among the top 50, used by over 60% including Qwen, sentence-transformers, and BERT. MIT is second, used by RoBERTa, BGE, and pyannote models. Some models like CLIP and Llama use custom licenses with varying commercial restrictions.

High downloads with low likes typically indicates the model is used programmatically as a dependency rather than being directly chosen. For example, google/electra-base-discriminator has 48M downloads but only 90 likes because it is pulled automatically by downstream packages. Models with high likes relative to downloads (like Kokoro-82M with 5,942 likes) tend to have strong community engagement.