Featherless

The first serverless inference platform for running any model on Hugging Face with zero cold starts.

Featherless eliminates the overhead of dedicated GPU provisioning by utilizing a multi-tenant architecture that keeps 25,000+ models warm and ready. Developers swap between Llama 3, Mistral, and specialized fine-tunes via a single OpenAI-compatible endpoint without managing instances or waiting for container boots. By decoupling model weights from active memory, the platform delivers sub-second time-to-first-token (TTFT) across the entire Hugging Face library. It is the leanest way to scale AI features: pay only for the tokens you generate while accessing a massive library of open-source intelligence on demand.

https://featherless.ai