AlphaGazette
Menu

AI

Hugging Face model routing via OpenRouter cuts capacity errors 46%

The handoff is already live for 160 models from 38 labs, and a request can move to another host if the first one fails. Developers’ own language-model keys are being withdrawn.

1 min read AI-generated

Hugging Face is sending language-model traffic through OpenRouter, and OpenRouter says the change cut capacity errors by 46 percent. The link is already live, at first for 160 models from 38 labs, so people chatting on a model page or calling those models should hit fewer dead requests when one host stumbles.

OpenRouter chooses the cheapest or the fastest outside host, it said. Hugging Face describes the choice as price, performance and reliability, and says a request that one host cannot finish is passed on automatically. New controls on model pages let people compare hosts on price, speed and uptime and inspect performance figures before they settle on one.

Hugging Face had steered this traffic itself since Inference Providers opened in January 2025, at first with Fal, Replicate, SambaNova and Together AI. A host in trouble then simply failed the call. The companies announced the joint setup on Monday. Co-founder and chief executive Clément Delangue called it “collaboration over competition,” with Hugging Face on storage and shared work and OpenRouter on inference.

Developer-supplied API keys for language models are being withdrawn, The Decoder reported. Both firms are also in the middle of acquisitions priced in the billions.