VRAM & GPU Cost Calculator

VRAM & GPU Cost Calculator

Estimates how much GPU memory a Hugging Face model needs to run or fine-tune, and shows the cheapest GPU setups with live rental prices. For people planning to host or train a model.

Made by fastgpu on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: VRAM & GPU Cost Calculator is a browser page that estimates the GPU memory a model needs. You pick a Hugging Face model, and it sizes memory for running, fine-tuning with LoRA, full fine-tuning, generating, transcribing or embedding. It then lists the cheapest GPU setups that fit and prices them live.

Who it suits: It suits people who plan to host or train a model and want to know what hardware to rent before they spend anything.

Industry | Who uses it | How they use it | Example
Machine learning | ML engineers | Check which GPUs can hold a model | Look up a model and read the cheapest setups
Startups | Founders | Estimate the cost of serving a model to users | Read the modelled cost per million tokens
Research | Researchers | Plan a fine-tuning run | Compare LoRA and full fine-tuning memory
Hobbyist AI | Enthusiasts | Find out if a model fits a home GPU | Try the 4-bit setting

Skip it if you need a measured benchmark, since the serving cost is a model and real stacks usually do worse.

Where to use it:

• A web browser, on the Space's page at Hugging Face, with no sign-in
• A local copy of the page files on Windows, macOS or Linux, opened from index.html

Use case | Who | What they get back
Size a model's memory | Engineers | GPU memory needed for the task and precision you choose
Find cheap hardware | Founders | GPU setups that hold the model, priced live with a link to each offer
Estimate serving cost | Operators | A modelled cost per 1M tokens, with the working shown

How to set it up: The online page needs nothing and asks for no sign-in. A local copy needs Python only to fetch the files.

• A web browser
• Python and pip, only to fetch a local copy
• Windows, macOS or Linux for the local copy

1. Install the Hugging Face hub library.
2. Download the Space's files into a folder named vram-and-gpu-cost-calculator.
3. Move into that folder.
4. Load index.html in a web browser.

pip install -U huggingface_hub
hf download fastgpu/vram-and-gpu-cost-calculator --repo-type space --local-dir vram-and-gpu-cost-calculator
cd vram-and-gpu-cost-calculator

How to use it: Pick a model in the search box, or paste its org and name. If your model is missing from the Hub, enter its size, for example 70B. Then read the memory figure, the GPU setups and the serving cost.

Tool | What it does
Search box | Lists thousands of popular public models, and falls back to the Hub's own search as you type
Precision | Chooses the number format for the estimate: as the model was published, or 16, 8 or 4 bits
Weights menu | For a GGUF repo, lists each build with its file size so memory is sized from the file you pick
Requests and context | Sets how many requests run at once and how long the context is, one request at 4K tokens by default

Example: You type 70B because the model is not on the Hub. The page shows how much GPU memory a model of that size needs, then lists the cheapest setups that hold it with their live prices.

For newcomers and experts: VRAM is the memory on a graphics card, and a model must fit in it to run well. Start by picking a model you know and read the cheapest setup the page lists.

The weights part of the estimate multiplies the parameter count by the bytes each one takes: 2 for 16-bit, 1 for 8-bit, half a byte for 4-bit. The KV cache and some headroom are added on top. A full fine-tune counts about 16 bytes per parameter plus activations, while LoRA counts frozen weights, adapters and optimizer state. A mixture of experts is sized on all its parameters. For GGUF, the estimate uses the size of the build file you pick. In practice such a build holds more than four bits for every parameter. The serving cost is modelled from memory bandwidth, so treat it as a floor.

Pick a model or type a size
Metadata is read from the Hub
Memory is sized for your task
Live GPU prices are matched
Cost per 1M tokens is modelled

Good to know:

• Coverage: The built-in list holds 14,192 well-downloaded public models. The list skips models small enough to need no rented GPU, and builds aimed at Apple silicon, CPUs or phones.
• Limit: The cost per 1M tokens is a model and not a measurement.
• License: mit.
• Cost to run: none named. The page needs no sign-in and uses a free match API.
• Data it reaches: Each model's metadata on the Hub, such as the parameter count and quantization config, and FastGPU's match API for sizing and live prices.

On Hugging Face, 13 November 2026

• Likes: 39
• Built with: static
• License: mit

Links

• Space: https://huggingface.co/spaces/fastgpu/vram-and-gpu-cost-calculator
• Files: https://huggingface.co/spaces/fastgpu/vram-and-gpu-cost-calculator/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.