Hugging Face Inference Providers Free Credits (2026)

Short answer: a free Hugging Face account no longer comes with free Inference Providers credits. To call hosted models through Hugging Face you either buy credits (pay-as-you-go at the provider's own price, no markup) or subscribe to PRO for $9/month, which includes $2.00 of compute credits every month. Team and Enterprise organizations get $2.00 per seat, pooled.

What Inference Providers is

Inference Providers is Hugging Face's serverless inference layer. Instead of deploying a model yourself, you call a model from the Hub and Hugging Face routes the request to one of its partner providers — currently including Cerebras, Groq, Together, Fireworks, Novita, Nscale, Replicate, fal, Featherless AI, DeepInfra, Baseten, Cohere, Scaleway, OVHcloud, Public AI, WaveSpeedAI, Z.ai and Hugging Face's own HF Inference. One Hugging Face token works for all of them, and billing lands on your Hugging Face account.

It replaced the old "Inference API (serverless)", which is now the hf-inference provider. As of July 2025, Hugging Face says hf-inference focuses mostly on CPU tasks such as embeddings, text ranking, text classification and small legacy models like BERT and GPT-2; most large LLMs and image models are served by the partner providers.

Who gets credits, and how much

The included credits are general-purpose compute credits. Besides Inference Providers, they can pay for Inference Endpoints, upgraded CPU and GPU hardware for Spaces (including ZeroGPU usage beyond your quota) and Jobs. They are added every month and applied automatically before any pay-as-you-go usage is billed. Hugging Face doesn't say unused credits roll over, so treat them as use-it-or-lose-it each month.

PRO also includes 8× the free ZeroGPU quota with the highest queue priority. If you mostly run demos in Spaces rather than API calls, that quota may be worth more to you than the $2 credit.

How billing works: routed vs your own key

Routed requests are charged at the same rate the provider charges, with no Hugging Face fee on top. For hf-inference, you pay compute time multiplied by the hardware's price. Hugging Face's own example: a FLUX.1-dev request that takes 10 seconds on a GPU costing $0.00012 per second is billed 10 × $0.00012 = $0.0012. At that rate, a PRO user's $2 would cover about $2 ÷ $0.0012 ≈ 1,660 such images in a month, if nothing else drew on the credits.

Step-by-step: getting set up

1. Create or sign in to your account at huggingface.co.
2. Decide how you'll pay: subscribe to PRO at huggingface.co/subscribe/pro for the $2 monthly credit, or add credits from your billing settings for pure pay-as-you-go. Free accounts must do one of these before routed requests will work.
3. Create a fine-grained access token in Settings → Access Tokens and enable Make calls to Inference Providers.
4. Optional: in Inference Providers settings, set your provider preference order or add custom provider keys.
5. Try models interactively in the Inference Playground (huggingface.co/playground), then call them from code.
6. Track spending on your billing page and the per-model, per-provider usage breakdown under Inference Providers settings.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://router.huggingface.co/v1",
    api_key=os.environ["HF_TOKEN"],
)
completion = client.chat.completions.create(
    model="openai/gpt-oss-120b:cheapest",  # or :fastest, :preferred, or :groq etc.
    messages=[{"role": "user", "content": "How many 'G's in 'huggingface'?"}],
)
print(completion.choices[0].message.content)

Choosing a provider to save credits

• :fastest (the default) picks the provider with the highest throughput in tokens per second.
• :cheapest picks the provider with the lowest price per output token — the obvious choice when you're spending a fixed $2.
• :preferred follows the order you set in your Inference Providers settings.
• Append a provider name (for example openai/gpt-oss-120b:groq) to force a specific host.
• GET https://router.huggingface.co/v1/models returns available models with per-provider pricing, context length, latency and throughput where available, so you can compare before you call.

The OpenAI-compatible router covers chat completions only. For text-to-image, embeddings, speech recognition and other tasks, use the huggingface_hub Python client or the @huggingface/inference JavaScript client, which handle provider-specific request formats for you.

Team and Enterprise: billing to the organization

Each member uses their own token, but requests only draw on the organization's pooled credits if you name the organization. In the Hugging Face clients pass bill_to="my-org-name" (Python) or billTo (JavaScript); with the OpenAI SDK or raw HTTP, send the header X-HF-Bill-To: my-org-name. Enterprise organizations with Resource Groups can pass a resource group ID instead to attribute costs. Admins can set a spending limit and disable specific providers in the organization settings, and Enterprise plans can pull a daily usage time series per member, model and provider through an API.

Forgetting the bill-to setting is the most common way Team members end up spending their personal credits (or getting payment errors on a free account) instead of the organization's pool.

Is PRO worth it just for inference?

Purely for API calls, usually not: PRO costs $9 a month and includes $2 of credits, so you're paying $7 for the other PRO benefits (ZeroGPU quota, priority, and so on). If you only need occasional API access, buying credits directly or using a provider's own free tier is cheaper. PRO makes sense if you also use Spaces with ZeroGPU, want one bill across many providers, or need Hub features that come with it.

What is still free on Hugging Face

Losing the free inference credit doesn't make Hugging Face itself paid. A free account can still browse and download open model weights from the Hub, run them on your own machine or server with libraries such as transformers, host Spaces on the free hardware tiers, including ZeroGPU within the free quota (PRO raises that quota eightfold). What now costs money is hosted, serverless inference through Inference Providers, because every routed request is executed and billed by a paid provider.

So if your goal is experimenting with an open model and you have a capable GPU or are happy with a small model on CPU, downloading the weights is the free route. If you need an HTTP endpoint without managing hardware, you'll either pay a little through Hugging Face or use one of the free tiers below.

Free alternatives if you just want API access

Several Inference Providers partners also run their own direct free tiers or credits; going to them directly is often the cheapest path when you only need one provider. See our guide to free AI API keys for a fuller list.

History of free inference on Hugging Face

• Before 2025: Hugging Face ran its own serverless Inference API, which is now the hf-inference provider.
• January 28, 2025: Hugging Face launched Inference Providers with fal, Replicate, SambaNova and Together AI as the first partners. At launch, signed-in free users got free inference with a small quota, PRO users got $2 of monthly credits, routed requests were billed at provider rates with no markup, and custom-key requests were billed by the provider.
• July 2025: hf-inference refocused mostly on CPU inference for embeddings, classification and small models.
• Since then: the partner list has grown to around 18 providers. The free tier was cut back: earlier versions of the pricing table showed a small monthly amount for free users (around $0.10), while the current table lists None and says free users need to purchase credits. Hugging Face hasn't published a dated announcement of that change.

Common mistakes

• Expecting a free account to work out of the box. Routed requests now need purchased credits or PRO.
• Using a classic read token. Create a fine-grained token with the Make calls to Inference Providers permission.
• Adding a custom provider key and expecting credits to apply. Requests made with your own provider key are billed by the provider, not from your monthly credits.
• Using the OpenAI-compatible router for images or embeddings. It handles chat completions only; use the Hugging Face clients for other tasks.
• Assuming the default provider is the cheapest. The default is :fastest; add :cheapest when cost matters.