OpenRouter Brings Multi-Provider Inference to Hugging Face

OpenRouter now powers chat model inference on Hugging Face Inference Providers, bringing its multi-provider routing, automatic failover and provider controls to developers building on the platform. The integration also lets OpenRouter users send request traces to Hugging Face Storage Buckets they own.
- Hugging Face Inference Providers now use OpenRouter routing for chat models.
- Routing considers live price, speed and reliability, with automatic failover.
- Users can prioritize fastest, cheapest or named providers.
- The integration covers 160 models from 38 labs.
- Request traces can be sent to a user-owned Hugging Face Storage Bucket.
OpenRouter routing comes to Hugging Face
Hugging Face Inference Providers now run chat models through OpenRouter, according to OpenRouter. Many models on Hugging Face are served by more than one provider, and OpenRouter selects a provider for each request using live price, speed and reliability data.
The setup does not require additional configuration for existing Hugging Face users. Their current Hugging Face token and code continue to work, while requests gain automatic failover. If a provider slows down or begins failing, requests can move to another provider. An error can also trigger a retry with the next provider during the same call.
OpenRouter said a live experiment conducted when it rebuilt the routing in October measured 46% fewer requests reaching a capacity error, 16.9% faster time to first token and 16.3% higher throughput, with 6.0% higher cost per request. These figures are from OpenRouter's experiment.
Users can choose how providers are selected
Developers can control the routing trade-off by adding a suffix to a model ID, such as :fastest or :cheapest, or by specifying a provider name. Requests work through OpenAI’s Chat Completions and Responses APIs or Anthropic’s Messages API.
OpenRouter said a conversation remains with the provider holding its prompt cache. Hugging Face model pages now include a Providers tab showing live price, speed and uptime for each provider. Users can also view spending by provider and disable providers they do not want to use.
The integration offers 160 models from 38 labs. The listed models include DeepSeek V4.1 Flash, MoonshotAI Kimi K3.1 Pro and Kimi K3, Z.ai GLM 5.3 and GLM 5.3 Flash, Qwen3.8 27B, MiniMax M3.1 Flash, Xiaomi MiMo-V2.6-Flash, NVIDIA Nemotron 3 Ultra and OpenAI gpt-oss-120b. OpenRouter said open-weight models account for about 70% of its paid tokens.
Tracing connects production traffic with Hub assets
OpenRouter can also send a trace of each request to a Hugging Face Storage Bucket owned by the user. The feature is in beta and is configured by adding Storage Buckets as a Broadcast destination in OpenRouter’s Observability settings.
Each JSON file contains the input messages, the model’s output, token counts and cost. Images and other media are excluded. Users can organize files by date or API key, creating source material for an evaluation set or a fine-tuning run on an open-weight model.
To begin, users need a Hugging Face token with the “Make calls to Inference Providers” permission and can open a supported model page on the Hub to chat or view code. Storage tracing requires a Hugging Face Storage Bucket, a token with Write permission, S3 credentials and the relevant namespace and bucket details in OpenRouter.
For users, the announcement combines broader model access with controls over provider selection and spending. It also places request data beside the models and datasets developers may use for evaluation or fine-tuning, according to OpenRouter.
OpenRouter: OpenRouter Now Powers Inference on Hugging Face


