01 What happened
Hugging Face has replaced its custom inference router for language models with OpenRouter's system, which dynamically chooses the best provider for each request based on real-time price, speed, and reliability and retries on another provider if the first attempt fails. The change took effect on November 16, 2026.
02 Key details
- Hugging Face switched Inference Providers for language models to OpenRouter's router, which selects the best provider per request by live price, speed, and reliability and retries on failure.
- Open-weight models accounted for 26.8% of tokens served by OpenRouter in October 2025 and 71.7% at the end of September 2026. In the weekly ranking cited by Hugging Face, DeepSeek V4.1 Flash 1017 led with 59.2 trillion tokens.
- According to OpenRouter, the new routing cut capacity errors by 46%, improved time to first token by 16.9% and throughput by 16.3%.
- Pricing is unchanged, with no markup from Hugging Face. Custom provider API keys for language models are no longer supported; image and video keys still work.
03 Why it matters
Automatic provider selection and retries reduce the work of managing access to open-weight models. The price remains the provider’s, with no markup from Hugging Face. Language model requests use the Hugging Face token.
04 Who it matters to
AI developers and enterprise AI customers.
Original source