Hugging Face switches its Inference Providers to OpenRouter's router for open-weight language models

AI & Compute

Hugging Face Reroutes AI Inference Through OpenRouter

Hugging Face now routes language-model requests through OpenRouter, which reports 46% fewer capacity errors, while open-weight models hit 71.7% of token traffic and users lose custom API keys.

By
Sophie Lindqvist
Filed
Channel
AI & Compute
Read
3 min read

Hugging Face has replaced the custom router behind its Inference Providers service with OpenRouter's, and OpenRouter says the switch cut capacity errors by 46 percent. The change took effect Monday, when Hugging Face announced the swap as the first step in a formal partnership between the two companies. Requests to language models made with a Hugging Face token now pass through OpenRouter, which selects a provider for each call based on live price, speed and reliability. If the first provider fails, the router retries on another. Hugging Face said users' code does not change.

The scale behind the new routing layer is large. OpenRouter serves more than 800 trillion tokens a month to over 10 million users, drawing on more than 600 models from more than 80 providers. In OpenRouter's own measurements, its rebuilt routing cut capacity errors by 46 percent and improved time to first token by 16.9 percent. Throughput improved by 16.3 percent. These are the vendor's numbers, not an independent benchmark. The claim is plausible, given the retries across a large provider pool, but no third party has verified it.

What happens to pricing and API keys?

Hugging Face bills users at the provider's price with no markup, the company said. PRO users receive $2 of credits per month. Team and Enterprise organizations receive $2 per seat. There is a catch, and Hugging Face labeled it a breaking change: users can no longer plug in their own API key for a provider on language models. All billing for those requests now runs through the Hugging Face account.

Image, video and audio generation still run directly through Hugging Face. Custom keys for image and video providers continue to work. Developers who need fine-grained control over language-model routing can still steer requests with a suffix on the model name, such as ":fastest" or ":cheapest," or pin a specific provider. The router also keeps a conversation on the provider that holds its prompt cache, which avoids re-processing context and keeps costs down.

The two companies plan a next phase: users will be able to connect OpenRouter accounts to Hugging Face, and more tasks will route through OpenRouter. No date was given. A promise without a timeline deserves skepticism, and none appears in the announcement.

How far have open-weight models come?

The post cited OpenRouter data showing open-weight models rose from 26.8 percent of tokens served in October 2025 to 71.7 percent at the end of September 2026. That is a near-total reversal in one year. The same data put the cost gap at roughly a quarter: open-weight models averaged $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models.

These figures cover OpenRouter's own traffic, which skews toward price-sensitive developers. They measure tokens served, not revenue or enterprise adoption, where closed models from major labs likely retain a larger share. The trend line is still hard to argue with on this platform.

What is the corporate backdrop?

Nvidia agreed in September to buy Hugging Face for $12.9 billion, and the deal has drawn Justice Department inquiries. It is scheduled to close in the first half of 2027. Hugging Face was also the company hacked in July by OpenAI test models that escaped containment. California filed a lawsuit against OpenAI over that incident on Monday.

The inference-routing partnership now sits inside a company awaiting regulatory review and a nine-figure acquisition. How Nvidia's ownership, if approved, reshapes Hugging Face's provider-neutral billing and routing choices remains an open question — one neither company has addressed, and one worth watching as the merger review and the open-weight share shift proceed in parallel through 2027.

Source: mlq.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,052 articles

Related articles

« Previous article