Hugging Face free ZeroGPU Spaces claim: what is confirmed so far

A claim is circulating that free Hugging Face accounts can now create two ZeroGPU Spaces. ZeroGPU is the company’s shared-GPU option for its hosted demo apps. Hugging Face has not confirmed the change in any company material reviewed for this article. Its most recent published policy, from September 2025, says only paying users can create ZeroGPU Spaces, while anyone can use them. This article sets out what Hugging Face has documented about ZeroGPU access, hardware and limits. It treats the free-account claim as unverified until the company confirms it.

What the free-account claim says and what Hugging Face has published

The claim is narrow: free accounts would be able to create two ZeroGPU Spaces. Until now, creating them has required a paid plan. Neither the number two nor any change to free-tier access appears in the Hugging Face sources available here. For now this is a user report, not a policy.

The clearest company statement on access comes from a September 2, 2025 Hugging Face blog post about speeding up ZeroGPU Spaces. It reads:

Pro users and Team / Enterprise org members can create ZeroGPU Spaces, while anyone can freely use them (Pro, Team and Enterprise users get 8x more ZeroGPU quota)

That sentence makes two separate distinctions. The first is about who can build on ZeroGPU: in the published policy, only paid individual or organization accounts. The second is about who can run other people’s ZeroGPU apps: anyone, with paid users getting a larger GPU-time allowance. The reported change would affect only the first.

A September 2025 post is more than a year old on today’s date. It is possible the policy has changed since without the change appearing in these sources. That is the reason for caution in both directions. The claim is not confirmed, and it is not disproven.

How ZeroGPU works

Spaces is Hugging Face’s platform for publishing machine-learning demo apps, often built with the Gradio interface library. On a standard GPU Space, the app grabs a GPU at startup and keeps it for the Space’s whole lifetime, even when no one is using it. Hugging Face says this is wasteful because demo traffic is “extremely sparse and spiky”.

ZeroGPU takes what the company calls a just-in-time approach. The main app process never attaches to the GPU’s CUDA software stack. Instead, ZeroGPU forks the process for each GPU task, attaches a GPU to the fork, runs the task, and then shuts the fork down to release the GPU. In Hugging Face’s description, that means:

• an app with no traffic uses no GPU;
• an app performing a task uses one GPU;
• an app can use multiple GPUs at once to handle tasks concurrently.

For developers, the code change is small. Hugging Face says the only change needed is to import its Python spaces package and add the @spaces.GPU decorator to the function that needs a GPU. The decorator intercepts PyTorch calls so that CUDA operations are postponed. The decorated function then runs in a fork when it is called.

When ZeroGPU first appeared, a Hugging Face staff member described it in a post on the platform as “a new hardware upgrade provided by HF where Spaces share GPUs or one Space can run on multiple GPUs”. That January 2024 thread also shows a different access model at the time. One user wrote that they had “requested to join the group” and modified their Spaces with the required decorators, hoping to be confirmed. Access has therefore changed at least once already. It moved from a group that users asked to join to the paid-plan rule stated in 2025.

ZeroGPU hardware and limits Hugging Face has documented

ZeroGPU runs on Nvidia H200 hardware. In its September 2025 post, Hugging Face says ZeroGPU “currently allocates” a MIG slice of an H200 using the 3g.71gb profile. MIG (Multi-Instance GPU) is a way of partitioning one physical GPU into several isolated slices. The company also said that more MIG sizes, including a full-slice 7g.141gb profile, “will come in late 2025”. The sources here do not show whether those larger slices launched as planned.

Item	What Hugging Face has stated
GPU	Nvidia H200
Default allocation	MIG slice, 3g.71gb profile
Planned sizes	Additional MIG sizes including full 7g.141gb slice, announced for late 2025
Who can create Spaces	PRO users and Team / Enterprise org members
Who can use Spaces	Anyone
Paid quota	8x more ZeroGPU quota for PRO, Team and Enterprise
Free accounts get two Spaces	Not confirmed by Hugging Face

The forking design has a cost that matters to anyone building on ZeroGPU. Because each GPU task usually gets a freshly started process, PyTorch’s just-in-time compiler, torch.compile, cannot reuse its work in memory. It has to restore compiled models from a filesystem cache instead. Hugging Face says this takes “from a few dozen seconds to a couple of minutes” depending on the model, which it calls “way too much for practical GPU tasks in Spaces”.

Hugging Face’s speed-up for ZeroGPU apps

The September 2025 post was mainly about a fix for that cold-start problem: PyTorch ahead-of-time (AoT) compilation. Compilation applies hardware-specific optimizations to a model, such as fusing operations together, to cut inference time. With AoT, a developer exports a compiled model once, saves it, and reloads it instantly in any process. Hugging Face says this is “exactly what we need for ZeroGPU”.

In Hugging Face’s own testing, AoT compilation gave speedups of 1.3x to 1.8x on models including Flux, Wan and LTX. Flux is an image model, and the post’s examples focus on image and video generation. The post also covers FP8 quantization, which stores numbers in a lower-precision 8-bit format to save memory and time. It also covers dynamic shapes, FlashAttention-3 and regional compilation. These are the maker’s own figures. No independent benchmark of ZeroGPU performance appears in the sources.

That work bears on the free-account question. If more people can create ZeroGPU Spaces, more apps compete for the same shared GPU slices. Techniques that shorten each task, such as AoT compilation, help the shared pool go further. Hugging Face has not linked the two publicly, and nothing in the sources says it plans to.

What a free ZeroGPU tier would mean for builders

Spaces sits inside a large platform. As of October 19, 2026, the Hub lists 3,167,979 models and 1,096,072 datasets, plus a large catalogue of Spaces, according to a pricing guide from software firm Fora Soft. The same guide says that, by Hugging Face’s spring 2026 count, the Hub serves more than 13 million AI builders. Spaces is a common way for model authors to put a working demo next to their weights.

Under the published rule, a developer who wants a shared H200 slice behind a demo needs a paid plan or a paid organization. If the reported change is real, a free account could host up to two GPU-backed demos without paying. That matters most for:

• Researchers and hobbyists who release open models and want a live demo next to the weights;
• Founders prototyping an image, video or language feature before committing to dedicated hardware;
• Educators and students who need GPU-backed examples but not an always-on server.

The limits would still apply. ZeroGPU is designed for demos with bursty traffic, not production workloads, and Hugging Face itself calls it “ideal for demos”. Paid users get 8x more quota than others. A free creator would likely face tighter GPU-time limits than a paying one, though no source states what any free-tier quota would be. Teams with steady traffic or strict latency needs have other Hugging Face options. Fora Soft’s guide lists managed Inference Endpoints and an Inference Providers API that routes to 18 partner companies. The guide frames dedicated GPUs as “a compliance, custom-weights or latency decision”.

Open questions

• Has Hugging Face changed who can create ZeroGPU Spaces since September 2025, and if so, when?
• If free accounts can create Spaces, is the cap two Spaces, and what GPU quota comes with them?
• Does the change apply to every free account, or only to some accounts or regions?
• Did the larger H200 slice sizes promised for late 2025 arrive, and are they open to all creators?

None of these questions can be answered from the sources available.

What to watch next

The deciding signal is a statement from Hugging Face itself. That could be an update to its ZeroGPU or pricing documentation, a blog post, or a staff post on the platform. Until one appears, the 2025 rule is the latest documented position: paid users create, anyone uses.

Developers who see the option on a free account should treat it as possibly limited or experimental. They should check the quota shown in their own Space settings rather than assume it matches paid tiers. Unicorn API will update this thread if Hugging Face confirms, changes or rules out free-account access to ZeroGPU. Related coverage of models published on the platform includes Nous Research’s Arete-27B on Hugging Face.