Hugging Face integrates OpenRouter directly into the Hub for open-weight model inference

The partnership addresses a common gap with open models: finding one is easy, but running it without provisioning GPUs is harder. With OpenRouter built into the Hub, developers use their existing Hugging Face token and a single endpoint to call open-weight models served by OpenRouter's provider network. Model pages now show live usage stats, and the inference widget has been redesigned with OpenRouter. Hugging Face's announcement post received 2,645 likes on X.
In practice, this turns the Hub from a model registry into something closer to an inference marketplace. OpenRouter already aggregates many providers and is where usage signals now come from: DeepSeek V4.1 Flash's 61.2T weekly tokens were measured there. Bringing that data onto model pages gives developers real adoption information next to the model card.
Competitively, it strengthens open-weight distribution just as several major releases move toward API-only access, including Alibaba's Qwen4-Max. CEO Clement Delangue also highlighted RL-adapted open models, such as a Qwen3.8-27B LoRA tuned for the pi coding agent that raised SWE-bench Pro Hard from 29.4% to 62.7%.
The timing is awkward in one respect. The same week, researchers hijacked the account behind the Kokoro TTS model on the Hub. More inference traffic running through Hub identifiers makes account and namespace integrity matter more. Watch for pricing transparency across providers and any provenance guarantees attached to served models.