Free Hugging Face Inference: Credits, ZeroGPU and Spaces

Hugging Face used to be the default answer to "free inference API". In 2026 the answer is narrower: free accounts get no monthly credits for Inference Providers, its serverless API. What stays free is GPU time on ZeroGPU Spaces, Spaces you call as an API, and routing requests through Hugging Face with your own key from a provider that has a free tier. Every number below was checked on Hugging Face's docs and pricing page on 2026-10-21.

What each account gets

Team ($20 per user a month) and Enterprise organizations get $2.00 per seat, shared by members (pricing, billing docs).

Inference Providers: no free credits

Inference Providers is one OpenAI-compatible endpoint, https://router.huggingface.co/v1, in front of partners such as Groq, Cerebras, Together, Fireworks, Novita, Scaleway and Hugging Face's own hf-inference. The billing page is plain: "Free users can use Inference Providers by purchasing credits, or by upgrading to PRO", and its table lists "None" as the monthly credit for free users. Hugging Face charges the provider's rate "with no additional fees".

PRO's $2.00 a month are general compute credits: they also pay for Inference Endpoints, Spaces hardware, ZeroGPU beyond the quota and Jobs. They apply only to requests routed through Hugging Face.

The free route through the same API: add a custom provider key in your Hugging Face settings. With your own key, "Hugging Face won't charge you for the call" and the provider bills you directly, so a provider's own free tier becomes your limit. Groq's free plan is one; the free Groq API key guide has its limits. Credits do not apply to custom-key requests.

ZeroGPU: free GPU minutes

ZeroGPU gives Gradio Spaces a shared NVIDIA RTX Pro 6000 Blackwell on demand: half a card (48 GB) by default, or a full card (96 GB) at twice the quota cost.

• Daily quota: 2 minutes unauthenticated, 5 minutes on a free account, 40 on PRO and Team, 60 on Enterprise. It "resets exactly 24 hours after your first GPU usage", and the quota you have left sets your queue priority.
• Default call length: 60 seconds of GPU per function call, raised with @spaces.GPU(duration=...).
• Hosting: free accounts in good standing can host 2 ZeroGPU Spaces (Gradio SDK only).

Kaggle's 30 GPU hours a week go much further for notebooks and training; the free GPU guide compares them, and the free GPU quota tracker adds Lightning AI and Paperspace.

Spaces: free hosting has changed

CPU Basic (2 vCPU, 16 GB RAM, 50 GB of non-persistent disk) still has no hourly cost, but "Gradio and Docker Spaces run on compute and require a paid plan to create". A free account can create Static Spaces, which are "free for everyone", plus its 2 ZeroGPU Gradio Spaces. Duplicating a Gradio or Docker Space follows the same rule (Spaces GPU upgrades). Spaces on free hardware "go to sleep" after a period without use (Spaces overview).

Steps: call a model for free

1. Sign up at huggingface.co and verify your email.
2. Create a fine-grained token on the Access Tokens page in your settings. For Inference Providers, tick Make calls to Inference Providers; for Spaces, read access is enough.
3. Call a ZeroGPU Space as an API. Every Gradio Space has a "Use via API" link in its footer. pip install gradio_client, then Client("owner/space", token="hf_...") and client.predict(..., api_name="/predict"). With your token, the call spends your account's 5 minutes instead of the 2-minute anonymous quota (Spaces as API endpoints).
4. Or route to a free provider with your own key. Add the provider key in your Inference Providers settings, then call curl https://router.huggingface.co/v1/chat/completions -H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" -d '{"model": "openai/gpt-oss-120b:groq", "messages": [{"role": "user", "content": "Hello"}]}'. The :groq suffix picks the provider; :fastest and :cheapest let Hugging Face choose.
5. Check your balance and usage on the billing page in your settings before you send real traffic.

Limits and the catch

• No free serverless credits. Without a custom key or purchased credits, routed Inference Providers calls are not free.
• Spaces are someone else's app. A public Space can change its API, sleep or be deleted; duplicate it to your account if you depend on it (Gradio duplicates need a ZeroGPU slot or a paid plan).
• Five GPU minutes go fast. A 45-second call on a full card spends 90 seconds of quota.
• Run models yourself instead: Hugging Face's Text Embeddings Inference serves open embedding models on your own hardware, and the free embeddings API guide lists hosted free tiers.