Hugging Face open models now run through OpenRouter
- Source
- x.com
- Date
Open models on @huggingface now run through OpenRouter. Each request goes to the provider with the best price, speed, and uptime for that model, with automatic failover. Everything runs on the same Hugging Face token and API. You can also send your traces to Hugging Face Storage with Broadcast, so your requests live on the Hub next to your models, ready for things like evals or fine-tuning.
Context
Hugging Face, where major models are published, already let developers call those models with one and one API through its Providers service, which forwards each request to a company that runs the model. On November 16 it replaced the router behind that service with OpenRouter's (Hugging Face's announcement). For language models, each request now goes to the provider that OpenRouter's live measurements of price, speed and reliability favor, and a request that fails is retried on the next provider within the same call. Hugging Face says existing code keeps working with the same token and that it bills the provider's price without adding a markup. That is what OpenRouter's post means by open models on Hugging Face now running through OpenRouter.
The switch covers language models only. Image, video and audio generation still go directly through Inference Providers. Because language-model requests are now billed to the Hugging Face account or organization, developers can no longer use their own API key for a particular provider with those models. They can still steer the router: adding :fastest or :cheapest to a model name selects a policy, a provider's name pins a single provider, and providers can be switched off in settings. For someone already calling open models through Hugging Face, the trade is automatic provider choice and failover in exchange for losing bring-your-own-key.
The second half of the post is about keeping records of requests, not about running models. Broadcast is an OpenRouter feature that sends a trace of each request, covering timing, cost, token counts and, unless privacy mode is on, the prompt and the response, to a destination the user picks. OpenRouter's setup guide shows how to make that destination a Hugging Face Storage Bucket, object storage that sits on the Hub next to a user's models and datasets, with each trace saved as its own JSON file. OpenRouter pitches this as a way to keep request data beside the models for evaluations or . The guide sets Broadcast up in OpenRouter's own settings, and Hugging Face's announcement does not mention it, so neither source says whether requests made with only a Hugging Face token can be traced this way.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Checking sign-in…
Loading comments…







