Hugging Face Now Routes Open-Weight Models Through OpenRouter
Hugging Face has switched its language model inference to OpenRouter, so users can call served open-weight models on the Hub with one Hugging Face token and automatic provider…

- Language model requests now automatically retry across providers for higher reliability while your existing code and Hugging Face token stay unchanged.
- Open-weight models reached 71.7 percent of tokens served on OpenRouter by September 2026, up from 26.8 percent in October 2025, and currently include the platform's most-used model.
- Pricing remains the provider rate with no markup; users can still select specific providers by suffix or disable any in settings.
What Changed
Hugging Face replaced its custom router for language models with OpenRouter's system. Every major open-weight model published on the Hugging Face Hub that a provider serves can now be called with your existing Hugging Face token through the same API.
OpenRouter picks the best provider for each request based on live measurements of price, speed and reliability across more than 80 providers. If the first provider fails, the request retries on another without any code changes. Image, video and audio generation continue to route directly through Hugging Face as before.
Benefits for Developers and Users
Developers and coding agents can use familiar tools and SDKs with setup guides provided for Claude Code, Codex, OpenCode, Pi and Hermes Agent. Model pages now show live provider lists with prices, speed and uptime plus metrics for the best combination.
Usage tracking breaks down spend by provider, and you can disable any provider in settings so the router never sends requests there. Conversation caching remains aware of which provider holds the cache; a stable session UUID can be sent in the X-HF-Session-Id header for consistency.
Usage and Performance Numbers
Open-weight models grew from 26.8 percent of tokens served on OpenRouter in October 2025 to 71.7 percent by the end of September 2026, with weekly volume rising more than 60-fold compared to nearly tenfold for closed models. This week the top model is the open-weight DeepSeek V4.1 Flash 1017.
Open-weight models cost about one-fourth as much on average. The best open-weight model scores lower than leading closed models on hard benchmarks but offers a strong price-performance trade-off for many workloads. OpenRouter's rebuilt routing improved time to first token by 16.9 percent, throughput by 16.3 percent and reduced capacity errors by 46 percent.
Pricing, Limits and Next Steps
You pay the underlying provider's price on your Hugging Face bill with no added markup. PRO users receive $2 monthly credits and Team or Enterprise plans get $2 per seat. Custom provider keys for language models are no longer supported, but they still work for image and video generation.
This is the first stage of the partnership. Future steps include connecting OpenRouter accounts to Hugging Face and routing more task types. Documentation and model pages have been updated with provider details and metrics.
Why it matters for uncensored AI
Our analysis starts from the source fact that open-weight models now account for over 70 percent of tokens on OpenRouter and include the current most-used model. Easier unified access through a single token and reliable routing can lower the barrier for developers and users who want to experiment with and build on these downloadable models.
This development gives people more practical choices when they prefer models that can be inspected, fine-tuned or run in environments they control. For those seeking greater user control over AI tools, the improved experience with open-weight models on Hugging Face expands options alongside fully self-hosted alternatives.
When you want an all-in-one hosted multimodal companion for uncensored chat, photos, voice and phone calls without managing infrastructure, we recommend considering Muah AI.


