Vibeleaderboard
← All Intel
Intel / article

Hugging Face and OpenRouter join forces to make open-weight models easy to run

Source
huggingface.co
Date

Context

Hugging Face, the site where most AI models are published for anyone to download, has handed the routing of its Inference Providers service to OpenRouter. Providers lets a developer call a model hosted on the Hub with a Hugging Face while an outside company actually runs it. Until now Hugging Face's own router picked one provider for each request and passed back that provider's error if the call failed. From November 16, 2026, requests to language models go through OpenRouter, a service that sends each request to one of more than 80 providers, and a failed attempt is retried on another provider inside the same call. Hugging Face says existing code keeps working unchanged.

OpenRouter picks the provider from live measurements of price, speed and reliability, and shifts traffic away from a provider that slows down or starts failing until it recovers. OpenRouter described that rebuilt router in October, when it moved away from ranking providers mostly by input price. Hugging Face repeats OpenRouter's measured results: 46% fewer capacity errors, a 16.9% improvement in time to first token and a 16.3% improvement in throughput. OpenRouter's own write-up also reported a 6.0% rise in cost, which the Hugging Face post leaves out. That follows from the design, since the default chooses the best mix rather than the lowest price. A developer who wants price alone can add a suffix to the model name, such as ":cheapest", or ":fastest" for throughput, or name a single provider. Routing also keeps a conversation on the provider that holds its prompt cache, the stored copy of earlier input that makes repeated cheaper to process.

Prices themselves do not change. Users pay the provider's rate on their Hugging Face bill with no Hugging Face markup, and PRO users and each seat on Team and Enterprise plans still get $2 of monthly credits. The one breaking change is that language-model requests are now billed through Hugging Face, so a developer can no longer use their own API key from a provider for these models. Providers a user or organization does not want can be switched off in settings, and the router will never send requests there. Image, video and audio generation stay on Hugging Face's existing system, outside OpenRouter, and custom keys for image and video providers keep working.

The router accepts the request formats of OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API for every chat model, which is what lets coding agents such as Claude Code or Codex point at Hub models directly. Hugging Face justifies the move with OpenRouter's usage data: open-weight models went from 26.8% of tokens served on OpenRouter in October 2025 to 71.7% at the end of September 2026, and cost about a quarter as much per token on average. Those shares come from OpenRouter's October analysis, which counted the top 50 models by weekly volume. The post also concedes that the best open-weight model still trails the best closed one, citing a score of 46 for Kimi K3.1 Pro against 65 for Claude Fable 5.2 on the Artificial Analysis Intelligence Index, at about a seventh of the cost per task.

For most existing users the main change is that a failed call is retried on another provider instead of returning an error, and model pages gain tabs showing which providers OpenRouter can route to, with their prices, speed and uptime. The people who need to act are those who relied on their own provider keys for language models, since that option is gone. Hugging Face calls this the first step of the partnership, with connecting OpenRouter accounts to Hugging Face and routing more tasks through OpenRouter listed as next.

Key takeaways · AI-distilled
  • Hugging Face Inference Providers now routes language-model calls through OpenRouter instead of its own router, picking a provider per request on live price, speed and reliability and retrying on another provider if the first fails. Existing code keeps working.
  • A suffix on the model id sets the routing policy: :fastest for throughput, :cheapest for price, or a provider name such as :novita to pin one. Routing is cache-aware, and an X-HF-Session-Id header keeps a conversation on the provider holding its prompt cache.
  • Breaking change: your own provider API keys no longer work for language models, because those requests are now billed through Hugging Face. Custom keys still work for image and video providers, which do not go through OpenRouter.
  • OpenRouter reports that its rebuilt routing cut capacity errors by 46% and improved time to first token by 16.9% and throughput by 16.3%. Previously, Inference Providers sent a request to one provider and returned its error if it failed.
  • Citing OpenRouter data, Hugging Face says open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026 and cost about a quarter as much on average, though the best open model still trails the best closed one on hard benchmarks.
Terms in this piece · Glossary
  • open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
  • inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Read the source huggingface.co
Recommended reads
Comments

Checking sign-in…

Loading comments…