Hot AI News

Hugging Face chat models now run on OpenRouter, and agent traces flow back

Clem Delangue announced the first part of a Hugging Face and OpenRouter partnership: OpenRouter now routes chat inference on Hugging Face, and traces can be stored in Hugging Face buckets.

Hugging Face chat models now run on OpenRouter, and agent traces flow back
Image: Hugging Face (Clem Delangue)

Clem Delangue, Hugging Face’s chief executive, said in a post on X that Hugging Face and OpenRouter are releasing the first part of a partnership. Chat models on Hugging Face Inference Providers now run through OpenRouter’s routing, and OpenRouter can store agent traces on Hugging Face.

Delangue summed up the split in one line: “Hugging Face for storage and collaboration. OpenRouter for inference.” He listed three results: running OpenRouter inference directly on Hugging Face, reaching open model weights from OpenRouter, and keeping OpenRouter agent traces on Hugging Face to inspect, train on or share.

The providers tab on a model page showing live price, speed and uptime per provider
Image: OpenRouter

Routing, failover and trace export in detail

OpenRouter published a companion post the same day with the specifics. Chat models on Inference Providers now pick a provider per request based on live price, speed and reliability. Existing Hugging Face tokens and code keep working. If a provider slows down or returns an error, the request moves to the next provider within the same call.

Developers can steer the choice with suffixes on the model ID, such as :fastest, :cheapest or a provider name. OpenRouter lists support for the OpenAI Chat Completions and Responses APIs and for Anthropic Messages. A conversation stays on the provider that holds its prompt cache. Each model page gets a providers tab showing live price, speed and uptime, and users can disable individual providers in their Hugging Face settings.

OpenRouter settings screen for sending traces to a Hugging Face Storage Bucket
Image: OpenRouter

The trace side is labeled beta. OpenRouter’s Broadcast feature can write each request as a JSON file to a Hugging Face Storage Bucket the user owns. According to OpenRouter, each file holds the input messages, the model output, token counts and cost, while images and other media are left out. Files can be organized by date or by API key. Setup means creating a bucket, generating S3 credentials from a write-enabled token, and switching on Broadcast under Observability in OpenRouter settings.

OpenRouter says the integration covers 160 models from 38 labs, meaning every model it serves that has a Hub page. Named examples include DeepSeek V4.1 Flash, Kimi K3.1 Pro, GLM 5.3, Qwen3.8 27B, Nvidia Nemotron 3 Ultra and OpenAI’s gpt-oss-120b.

Advertisements

What changes for developers and teams

For developers already calling the Hugging Face router, the practical change is underneath: the same token now gets OpenRouter’s failover and caching behavior. OpenRouter’s own figures, from a live experiment after its October routing rebuild, show 46% fewer capacity errors, a 16.9% faster time to first token and 16.3% higher throughput. The same experiment showed a 6.0% higher cost per request. Those are the company’s numbers, and the post does not say how they apply to Hugging Face traffic specifically.

The market context explains the pairing. OpenRouter says open-weight models account for about 70% of the tokens its users pay for. Hugging Face’s documentation now states that routing is powered by OpenRouter and that provider prices carry no markup. Hosting is spread across many providers, so a router that can shift traffic matters. A model such as GLM-5.3, which needs two DGX Stations to run locally, is far easier to try through an API than on owned hardware.

The trace export targets teams building agents. Logged inputs, outputs and costs are the raw material for evaluation sets and fine-tuning data, and OpenRouter names both uses. Storing them in a bucket the team owns keeps that data outside a third-party dashboard. Anyone sharing traces publicly should check first what the prompts contain, since the files hold full message text.

What the announcements leave unclear

Delangue calls this the first part, and neither source describes the rest. His claim that open model weights can be reached directly from OpenRouter has no matching detail in OpenRouter’s post, which says nothing about weight files, sizes or downloads. Neither source gives per-token prices beyond pointing to the live providers tab, and neither states whether the trace export has plan limits or extra storage fees. The performance figures are OpenRouter’s own and have not been independently verified. Readers who want to judge what a model’s listing actually promises can start with our guide to reading model cards.

Access is available now: create a Hugging Face token with the permission to make calls to Inference Providers, open a supported model’s Hub page, and use the code snippets there. Hugging Face’s documentation says PRO, Team and Enterprise plans include monthly credits for inference, and pay-as-you-go applies otherwise.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like