Hugging Face Free API 2026: the Free Tier Is Now $0.10/mo

Last updated: 11 September 2026. Checked against Hugging Face's docs, pricing pages and live router API that day.

The Hugging Face Inference API free tier in 2026 is $0.10 of credit a month, not a request limit. Free accounts spend it at partner providers' own prices through Inference Providers. When it runs out, any user can buy more credits by card; PRO ($9 a month) includes $2. The old "few hundred requests an hour" scheme is gone: Hugging Face's rate-limit page for inference now redirects to the pricing page.

Three more things most guides get wrong:

Hugging Face's own serverless service, hf-inference, has run mostly on CPU since July 2025. It serves classification, embeddings and similar tasks, one image model and no chat models.
The old api-inference.huggingface.co address no longer resolves. Chat calls go to router.huggingface.co/v1.
Creating Gradio or Docker Spaces now needs a paid plan, although free accounts can still host two ZeroGPU Spaces.

Free vs PRO vs Team vs Enterprise

	Free	PRO	Team	Enterprise
Price	$0	$9/month	$20/user/month	$50/user/month (or custom)
Inference credits	$0.10/month	$2/month	$2 per seat, shared	$2 per seat, shared
After credits run out	Buy credits	Buy credits	Buy credits	Buy credits
ZeroGPU quota per day	5 minutes	40 minutes	40 minutes	60 minutes
Hub API requests per 5 minutes	1,000	2,500	3,000	6,000
Spaces you can create	Static, plus 2 ZeroGPU Gradio Spaces	Gradio, Docker, 10 ZeroGPU	Up to 50 ZeroGPU	Up to 50 ZeroGPU

Sources: Hugging Face's Inference Providers pricing, Hub rate limits, ZeroGPU docs and pricing page, 11 September 2026. Hugging Face marks the free credit "subject to change", and says the free and anonymous rate limits change "depending on platform health".

What $0.10 a month actually buys

Inference Providers bill at each provider's price per million tokens, with "no extra markup". Prices from Hugging Face's live router model list on 11 September 2026:

Model	Cheapest provider on HF	Output price per 1M tokens	Output tokens for $0.10
Llama 3.1 8B Instruct	Novita / DeepInfra	$0.05	~2,000,000
gpt-oss-20b	DeepInfra	$0.14	~700,000
gpt-oss-120b	DeepInfra	$0.17	~590,000
gpt-oss-120b	Groq	$0.75	~130,000
GLM-4.7-Flash	DeepInfra	$0.40	~250,000
Kimi K2.5	DeepInfra	$2.25	~44,000
DeepSeek V4 Pro	DeepInfra	$2.60	~38,000

Output tokens only; input tokens cost extra, usually far less. So the free credit is enough to check that a model works in your code, and roughly a day of light testing at best. It is not a way to run anything.

The router list also shows a price of $0 for a handful of model and provider pairs, such as Qwen 3.8 27B on OVHcloud and inclusionAI's Ling 3.0 Flash models on Novita, while its separate is_free promo flag was off for every model. We could not confirm without a paid call whether those calls really cost nothing.

Rate limits on the free tier

No per-hour or per-minute request limit is documented for Inference Providers any more; the credit is the limit. What remains are the Hub's own limits, counted over 5-minute windows:

Plan	Hub API calls	File downloads (resolvers)	Page views
Anonymous (per IP)	500	3,000	100
Free	1,000	5,000	200
PRO	2,500	12,000	400
Team	3,000	20,000	400
Enterprise	6,000	50,000	600

Going over returns HTTP 429 Too Many Requests. Hugging Face does not say whether inference calls through the router count toward these, and does not document the error you get when the monthly credit is spent; it recommends turning on automatic recharge "to avoid service disruptions after credits are exhausted".

hf-inference: Hugging Face's own serverless API, now mostly CPU

"HF Inference is the serverless Inference API powered by Hugging Face. This service used to be called 'Inference API (serverless)' prior to Inference Providers." Since July 2025 it "focuses mostly on CPU inference (e.g. embedding, text-ranking, text-classification, or smaller LLMs that have historical importance like BERT or GPT-2)".

What it serves: about 1,460 models in the live Hub API, mostly text classification (287), sentence similarity (285), fill-mask (259), feature extraction (143) and image classification (140). One text-to-image model (Stable Diffusion 3 Medium) and no text-generation models.
How it bills: after the free credit, "compute time x price of the underlying hardware". Hugging Face's example: a 10-second FLUX.1-dev request on a GPU costing $0.00012 a second is billed $0.0012.
So does the free inference API always use GPU hardware? No. hf-inference runs mostly on CPU; GPU work, including LLMs and image generation, goes to the partner providers.

If you are following an old tutorial that calls https://api-inference.huggingface.co/models/...: that host no longer has a DNS record. Use the router or the huggingface_hub library instead.

Image and video generation

Task	Providers with models	Notes
Text-to-image	fal, Replicate, WaveSpeedAI	1,000+ models each, LoRAs included
Text-to-image	Nscale, hf-inference	One model each (FLUX.1-schnell, SD3 Medium)
Text-to-video	WaveSpeedAI (191), fal (32), Replicate (3)	Model counts from the live Hub API

Hugging Face's partner table also lists Together and Novita for image or video, but the live API returned no such models for either. Image and video are billed at the provider's own rates, which Hugging Face does not republish; check the model page before running a batch. At the hf-inference example rate above, $0.10 covers about 80 ten-second GPU generations, and partner prices differ.

Video in particular is not something $0.10 goes far on. For regular image work, the practical free option on Hugging Face is a ZeroGPU Space (see below), not the API.

Inference Providers: the router and its 18 partners

One OpenAI-compatible endpoint for chat models:

from openai import OpenAI

client = OpenAI(base_url="https://router.huggingface.co/v1", api_key="hf_...")
reply = client.chat.completions.create(
    model="openai/gpt-oss-120b:cheapest",
    messages=[{"role": "user", "content": "Hello"}],
)

The suffix picks the provider: :fastest (the default, highest tokens per second), :cheapest (lowest output price), :preferred (your order in settings), or a provider name such as openai/gpt-oss-120b:groq. The router is "currently available for chat completion tasks only"; images, audio and embeddings go through the huggingface_hub client. Credits do not apply if you plug in your own provider key.

Partners on 11 September 2026: Baseten, Cerebras, Cohere, DeepInfra, fal, Featherless AI, Fireworks, Groq, HF Inference, Novita, Nscale, OVHcloud, Public AI, Replicate, Scaleway, Together, WaveSpeedAI and Z.ai. DeepInfra joined in April 2026 and Baseten in August; Nebius, SambaNova and Hyperbolic are no longer listed. The router's model list had 138 chat models across 14 providers.

The router also reports each provider's throughput. For gpt-oss-120b on 11 September:

Provider	Throughput	Price in / out per 1M
Cerebras	1,097 tok/s	$0.35 / $0.75
Groq	427 tok/s	$0.15 / $0.75
Baseten	159 tok/s	$0.10 / $0.50
Together	127 tok/s	$0.15 / $0.60
Nscale	107 tok/s	$0.10 / $0.40
Fireworks	93 tok/s	$0.15 / $0.60
Novita	79 tok/s	$0.05 / $0.25
DeepInfra	32 tok/s	$0.037 / $0.17

The fastest provider costs about four times the cheapest per output token. :cheapest is the right default for tests on the free credit; :fastest for a demo someone is watching.

Inference Endpoints pricing (dedicated)

For a model the providers do not serve, such as your own fine-tune, Inference Endpoints runs it on dedicated hardware. Prices are shown per hour and billed per minute, and you need a paid subscription and credits on the account.

Instance (AWS)	1 GPU	Larger
CPU (Intel Sapphire Rapids)	$0.033/h	up to $0.536/h (x16)
NVIDIA T4	$0.50/h	$3/h (x4)
NVIDIA L4	$0.80/h	$3.80/h (x4)
NVIDIA A10G	$1.00/h	$5/h (x4)
NVIDIA L40S	$1.80/h	$8.30/h (x4), $23.50/h (x8)
NVIDIA A100	$2.50/h	$20/h (x8)
NVIDIA H200	$5.00/h	$40/h (x8)
NVIDIA RTX PRO 6000	$2.75/h	$22/h (x8)

H100 ($4.50) and B200 ($9.25) are marked deprecated since December 2025 in the docs, though the pricing page still lists them. On GCP, T4 is $0.50, L4 $0.70 and A100 $3.60 an hour. Source: Endpoints pricing.

Scale to zero: an endpoint can go idle after a period of inactivity (1 hour by default) and stops costing money. The first request after that gets HTTP 503 while it wakes, unless you send the X-Scale-Up-Timeout header to wait. For bursty traffic this is what keeps a dedicated endpoint affordable.

Spaces free limits and ZeroGPU

CPU Basic hardware is free: 2 vCPU, 16 GB RAM and 50 GB of disk that is not persistent. A Space sleeps after 48 hours without visitors.
Creating Spaces changed: "Static Spaces are free for everyone. Gradio and Docker Spaces run on compute and require a paid plan to create". "Free personal accounts in good standing can still host up to 2 Gradio Spaces running on ZeroGPU" (verified email, account older than 30 days). PRO users can host 10.
ZeroGPU now runs on NVIDIA RTX Pro 6000 Blackwell cards: large is half a card with 48 GB, xlarge a full card with 96 GB at twice the quota.

Account	ZeroGPU per day
Not signed in	2 minutes
Free	5 minutes
PRO	40 minutes (extensible)
Team	40 minutes
Enterprise	60 minutes

The quota "resets exactly 24 hours after your first GPU usage". Since April 2026, PRO, Team and Enterprise users can go over it at $1 per 10 minutes.

Is PRO worth $9?

For inference alone, not really: PRO turns $0.10 of credit into $2, and $1.90 of pay-as-you-go credit costs less than $9. PRO pays off if you use Spaces: it lets you create Gradio and Docker Spaces, gives 8 times the free ZeroGPU quota (40 minutes a day on a Blackwell GPU) with priority in the queue, and 1 TB of private storage. Its credits also work on Endpoints, upgraded Spaces hardware and Jobs.

Hugging Face vs OpenRouter

Both give one OpenAI-compatible API over many open models. OpenRouter has a real free tier: 19 free models at 50 requests a day, or 1,000 once you have bought $10 of credits. Hugging Face has no free models, only the $0.10 credit, but passes provider prices through with no markup and adds the Hub: models, datasets, Spaces and Endpoints on one account. For free testing of chat models, OpenRouter's free models go further; for anything tied to the Hub, or image and video models, Hugging Face is the natural choice.

What changed since this page's June version

June version said	September 2026
Serverless free tier: a few hundred requests/hour, models under 10B	$0.10/month of credit; no request limit documented
Call `api-inference.huggingface.co/v1`	That host is gone; use `router.huggingface.co/v1`
PRO: "2M monthly Inference Provider credits"	$2 of credit a month
PRO: 25 minutes of daily H200 ZeroGPU	40 minutes a day on RTX Pro 6000 Blackwell
Endpoints: H100 $6/h, A100 $4.50/h, A10G $1.30/h	H100 deprecated ($4.50), A100 $2.50, A10G $1.00, H200 $5.00
Partners include Nebius, SambaNova, Hyperbolic	Replaced; DeepInfra and Baseten joined in 2026
(not covered)	Gradio and Docker Spaces need a paid plan to create

Frequently asked questions

What are the Hugging Face Inference API free tier limits in 2026? Free accounts get $0.10 of Inference Providers credit a month, spent at partner providers' own prices; Hugging Face no longer documents a request-per-hour limit for inference. When the credit is gone you can buy more by card, or subscribe to PRO for $9 a month, which includes $2 of credit. Separately, the Hub allows free accounts 1,000 API calls per 5 minutes.

Is the Hugging Face Inference API free? Only up to $0.10 a month. That is enough for a few hundred thousand output tokens on a cheap model such as gpt-oss-120b at the lowest-priced partner, or a few dozen thousand on a large model such as DeepSeek V4 Pro. After that you pay the provider's price, with no Hugging Face markup.

Does the free Hugging Face inference API use GPU hardware? Not by default. Hugging Face's own hf-inference service has focused mostly on CPU since July 2025 and serves tasks like embeddings, classification and text ranking. LLMs and image or video generation run on partner providers' GPUs and are paid from your credit.

Can I generate images or video for free on Hugging Face? Within the $0.10 monthly credit, a little: text-to-image models are served by fal, Replicate, WaveSpeedAI, Nscale and hf-inference, and text-to-video by WaveSpeedAI, fal and Replicate, each at its own price. For regular free image generation, a ZeroGPU Space with 5 minutes a day of GPU time on a free account is the more practical route.

What happened to api-inference.huggingface.co? The old serverless Inference API became hf-inference, one of the providers behind Inference Providers, and the api-inference.huggingface.co host no longer resolves. Chat models are called through router.huggingface.co/v1, which is OpenAI-compatible; other tasks go through the huggingface_hub client.

How much do Hugging Face Inference Endpoints cost? On AWS, from $0.033 an hour for a CPU instance, $0.50 for an NVIDIA T4, $0.80 for an L4, $1.80 for an L40S, $2.50 for an A100 and $5.00 for an H200, billed per minute. Endpoints can scale to zero after an idle period so you stop paying when there is no traffic.

What are the Hugging Face Spaces free limits? Free CPU Basic hardware has 2 vCPU, 16 GB of RAM and 50 GB of non-persistent disk, and sleeps after 48 hours without visitors. Free accounts can create static Spaces and up to two ZeroGPU Gradio Spaces, with 5 minutes of ZeroGPU time a day; creating other Gradio or Docker Spaces needs a paid plan.