Skip to main content
Back to timeline
OpenRouter BlogSource publication:

OpenRouter's routing now drives Hugging Face inference: 160 models pick price, speed and reliability per request

Synopsis

Chat-model inference on Hugging Face now runs through OpenRouter's provider routing, which picks a provider for each request from live price, speed and reliability and fails over automatically. Previously developers had to choose among providers themselves. OpenRouter reports that a live experiment when it rebuilt this routing in October measured 46% fewer requests hitting a capacity error, 16.9% faster time to first token and 16.3% higher throughput, at 6.0% higher cost per request. 160 models from 38 labs are available this way today.

Source-provided article image: OpenRouter Now Powers Inference on Hugging Face
OpenRouter

Interpretation

When calling a chat model on Hugging Face, provider selection is no longer a manual developer decision: OpenRouter's provider routing makes it per request from live price, speed and reliability, and existing Hugging Face tokens and code keep working. Previously a single Hugging Face chat model could be served by several providers, and weighing which one to use fell to the developer. OpenRouter reports that a live experiment when it rebuilt this routing in October measured 46% fewer requests hitting a capacity error, 16.9% faster time to first token and 16.3% higher throughput, for 6.0% higher cost per request, and says builders on Hugging Face now get the same routing.

The routing also brings automatic failover: if a provider slows down or starts failing, requests move to other providers, and an error is retried on the next provider inside the same call. Developers can steer priorities with a model-ID suffix such as :fastest, :cheapest or a provider name, and requests work through OpenAI's Chat Completions and Responses APIs or Anthropic's Messages API. A conversation stays on the provider holding its prompt cache; each model page gains a Providers tab with live price, speed and uptime per provider, plus spend by provider and the ability to turn off providers.

The integration runs both ways: OpenRouter can write a trace of each request to a user-owned Hugging Face Storage Bucket, with each file holding input messages, model output, token counts and cost, while images and other media are left out. Those files are raw material for building an evaluation set or a fine-tuning run on an open-weight model, and can be organized by date or API key. The Broadcast destination is in beta; 160 models from 38 labs are available this way today, every model OpenRouter serves that has a page on the Hub, including DeepSeek V4.1 Flash, Kimi K3.1 Pro and K3, GLM 5.3 and 5.3 Flash, Qwen3.8 27B, MiniMax M3.1 Flash, MiMo-V2.6-Flash, Nemotron 3 Ultra and gpt-oss-120b.

Perspective

The demonstrated scope is chat-model inference routing on Hugging Face Inference Providers and writing request traces to a user-owned Storage Bucket; trace writing is labeled beta and excludes images and other media.

The routing performance figures come from one live experiment OpenRouter ran when rebuilding the routing in October, and the material does not state the models, traffic scale or statistics used; trace writing remains in beta.