Hugging Face switched its Inference Providers language-model traffic to OpenRouter's router on November 16, so one Hugging Face token now reaches every provider OpenRouter routes to. The stakes: open-weight models already carried 71.7% of OpenRouter's tokens at the end of September, and the switch ends bring-your-own provider keys for language models.
Key facts
- Hugging Face switched Inference Providers language-model calls to OpenRouter's router on November 16, 2026; image, video and audio generation still bypass OpenRouter.
- Open-weight models went from 26.8% of OpenRouter tokens in October 2025 to 71.7% at the end of September 2026, peaking at 77.5% in July.
- DeepSeek V4.1 Flash 1017 is the most-served model this week at 59.2 trillion tokens per the announcement blog; OpenRouter's live rankings show 61.2T through the Nov 16 bucket.
- The switch is breaking for BYOK: custom provider API keys no longer work for language-model calls, though they keep working for image and video providers.
- OpenRouter reports more than 800 trillion tokens a month served to over 10 million users across 600+ models from 80+ providers.
- Hugging Face charges no markup on provider pricing; PRO accounts get $2 monthly credits and Team/Enterprise get $2 per seat.
What happened
Before the switch, Hugging Face's Inference Providers picked one backend for your request and returned that backend's error if it failed. Now the request goes through OpenRouter, which selects a provider per call using live measurements of price, speed and reliability, and retries on the next-best provider within the same call when the first one fails or rate-limits.
Routing is cache-aware, so an ongoing conversation stays on the provider holding its prompt cache; Hugging Face documents an X-HF-Session-Id header to make session pinning explicit. Developers can still steer the router with a suffix on the model ID: :fastest for highest throughput, :cheapest for lowest price, or :provider (for example zai-org/GLM-5.3:novita) to pin one backend. Turning a provider off in Inference Providers settings removes it from the router's candidate list, per-user or per-organization.
OpenRouter's own routing got rebuilt recently, and the changes Hugging Face is inheriting are measurable on OpenRouter's terms: the company reports the new router cut capacity errors by 46%, improved time to first token by 16.9% and throughput by 16.3%. Those are OpenRouter's numbers, not an independent audit, but the mechanism, routing on live traffic rather than list price, is visible in the product.
The pricing model does not change. You pay the provider's price on your Hugging Face bill, and Hugging Face adds no markup. What does change is keys: language-model requests are billed to your Hugging Face account or organization through the OpenRouter path, so your own API key for a specific provider no longer applies to those calls. Model pages also gained Providers and Metrics tabs showing per-provider price, speed and 24-hour uptime, and the usage overview now breaks spend down by the provider each request actually went to.
Why it matters
For a builder, the practical effect is that a single Hugging Face token becomes a live auction across every provider serving a model. You no longer pick a vendor; you pick a routing policy, and the router picks the vendor per request. That compresses the value of raw price differences between hosted providers, because the router arbitrages them away by default. The :cheapest suffix exists precisely because ordinary routing will not always pick the lowest price.
It also shifts the differentiation question. When any provider on the router can be swapped in per request, what remains sticky is what the router cannot arbitrage: Zero Data Retention terms, license conditions, regional hosting and enterprise billing. That is the same conclusion our zero data retention coverage keeps reaching. Retention policy, not tokens per second, is where provider deals are won.
The usage backdrop gives the switch its timing. Open-weight share on OpenRouter climbed from 26.8% in October 2025 to 71.7% eleven months later, and weekly open-weight volume grew more than 60-fold in that window against roughly tenfold for closed models. We treated the earlier open-model token surge as a demand signal; wiring the largest open-model distribution channel into the largest open-model router is the supply-side response to it. Pricing pressure across providers, tracked in last week's price-index delta, gets sharper when the default path routes around list prices automatically.
Background
Hugging Face launched Inference Providers as a curated set of hosted backends callable with a Hugging Face token, and it ran its own router until now. OpenRouter, for its part, says it serves more than 800 trillion tokens a month to over 10 million users across 600+ models from 80+ providers, routing on live measurements of every provider across that traffic. The two are now one funnel: every open-weight model on the Hugging Face Hub that any OpenRouter provider serves is callable through the Hub's API surface.
The switch also matters for the model ecosystem, not just the plumbing. The most-served model on OpenRouter this week is an open-weight one, DeepSeek V4.1 Flash 1017 at 59.2 trillion tokens per the announcement blog, and the hub's example models (Inkling 2 Small, Kimi K3.1 Pro, Qwen3.8 27B) are all open weights. Distribution through Hugging Face's default developer path now runs through the same rankings that decide which open models get adopted. For how the major hosted options compare on price and retention outside this partnership, see our MiniMax M3.1 Flash API rundown and the provider price index.
One caveat belongs in any honest read of the adoption numbers: these are OpenRouter's own measured router statistics, reasonably verifiable against its public rankings, not an independent market survey. A provider that sells most capacity through its own API will look small on this meter. The 71.7% figure measures one large router, not the enterprise market.
What's next
Hugging Face names the next steps in the announcement: connecting your OpenRouter account to Hugging Face, and routing more task types through OpenRouter. Developers with pinned providers should check the :provider suffix path and their settings before the next deploy, because the BYOK removal is live today. The number to watch is whether open-weight share holds above the July peak of 77.5% as closed-model refreshes land; OpenRouter's rankings page updates daily and its data is CC BY 4.0 licensed, so the trend is independently chartable.
Questions
- What changed with Hugging Face Inference Providers on November 16?
- Requests to language models now route through OpenRouter instead of Hugging Face's own router. OpenRouter picks a provider per request based on live price, speed and reliability, retries on another provider inside the same call, and keeps conversations on the provider holding their prompt cache. Image, video and audio calls are unchanged.
- Can I still use my own API key for a provider on Hugging Face?
- Not for language models. Because LM requests now flow through OpenRouter and are billed to your Hugging Face account or organization, custom provider keys no longer work for them. You can still pin a provider with a :provider suffix or disable providers in settings. Custom keys keep working for image and video providers.
- Does Hugging Face add a markup on OpenRouter-routed calls?
- No. Per Hugging Face's announcement, you pay the provider's price on your Hugging Face bill and Hugging Face adds no markup. PRO accounts get $2 in credits monthly, and Team and Enterprise organizations get $2 per seat.
- What share of OpenRouter traffic is open-weight models?
- Per OpenRouter data cited by Hugging Face, open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, after peaking at 77.5% in July. That is one router's measured traffic, not a market-wide survey.
Sources
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
