tech

The hottest AI models aren’t the ones developers actually use

Silicon Valley goes crazy for every new frontier model. However, new data from Hugging Face shows that developers gravitate to lesser known models.

The hottest AI models aren’t the ones developers actually use

TL;DR

  • Hugging Face data shows a disconnect between hyped frontier AI models and those developers actually use.
  • Developers prefer smaller, older models that are cheap, stable, and already in production.
  • All-MiniLM-L6-v2, a small 2021 model, had 1.55 billion downloads in the first 7 months of 2026 with only 5,156 likes.
  • Models under 1 billion parameters account for 83% of all-time downloads on Hugging Face.
  • Large models (over 100 billion parameters) represent only 1% of all-time downloads.
  • Chinese frontier models like Moonshot AI's Kimi K3, despite significant hype, show low download-to-like ratios.
  • Alibaba's Qwen model series, with a range of sizes, has achieved high usage due to its integration into developers' workflows.
  • Companies like Pinterest adopt a model-agnostic approach, using different types of AI based on cost, performance, and economics.