Hugging Face unifies serverless inference with SambaNova, Replicate, and more

Hugging Face is finally opening up its serverless Inference API to third-party providers, a move that acknowledges a simple truth: they’re better at hosting models and building community than running fast, scalable inference themselves. Announced today, the new Inference Providers feature lets developers route requests through partners like SambaNova, Replicate, Together AI, and fal.ai directly from the Hub, using either their own API keys or Hugging Face’s billing system.

The timing is perfect. With models like DeepSeek-R1 achieving mainstream fame practically overnight, the demand for fast, reliable inference on open models has never been higher. The integration is refreshingly straightforward. In the Hub’s UI, you can set API keys for your preferred providers and rank them. On a model page, compatible providers are sorted by your preference, so you’re not hunting around for who serves what. For developers, the SDK experience is just as clean: swap a single provider parameter in your InferenceClient call and everything else stays the same.

Rodrigo Liang, CEO of SambaNova, framed the partnership around speed, noting developers will get “much faster inference speeds on a wide range of the best open source models.” Replicate’s founding designer Zeke Sikelianos called Hugging Face “the de facto home of open-source model weights” and revealed that Replicate already uses the Hub as its internal weights registry. That’s a telling detail — the companies powering inference were already dependent on Hugging Face for model distribution. This formalizes that relationship.

What’s smart here is the routing layer. You can use the proxy URL https://router.huggingface.co/{provider} as a drop-in base URL for OpenAI-compatible clients. That means tools and applications already built around the OpenAI API format can switch to open models on any supported provider by changing one URL and a model name. I suspect this is the real play: Hugging Face becomes the neutral switchboard for serverless inference, while providers compete on speed and price underneath. Whether that competition actually drives down costs or just adds abstraction remains to be seen.