LLM VRAM Calculator

LLM VRAM Calculator

Estimates how much GPU memory a Hugging Face model needs at any quantization and context length, split into weights, KV cache and overhead, and shows which GPU sizes fit. It suits anyone planning to run a local LLM.

Made by donaldJJ 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: This is a static web page that estimates the GPU memory a language model needs. You enter a Hugging Face model id, and the page reads the model's configuration and parameter count in your browser. It then shows memory use for the quantization and context length you choose.

Who it suits: It suits people who run language models on their own GPUs and want to know before downloading whether a model will fit.

Industry | Who uses it | How they use it | Example
Local AI | Hobbyists with one GPU | Check whether a model fits at a chosen quantization | A hobbyist tests a 27B model on a 24 GB card.
ML engineering | Engineers planning deployments | Estimate memory for weights, cache and overhead | An engineer compares context lengths before sizing a server.
Model publishing | Model authors | Add a VRAM badge to a model card | An author links the badge to their own model.
Research | Students | Learn how KV cache grows with context | A student changes the context length and watches the cache figure.

If you need an exact measured peak, skip it, because the page gives planning estimates.

Where to use it:

• A web browser, through the Space's page on Hugging Face
• Any browser on Windows, macOS or Linux, opening a downloaded copy of index.html

Use case | Who | What they get back
Size a GPU purchase | Local AI users | The GPU sizes that fit, from 8 GB up to 96 GB
Plan context length | Engineers | The longest context each GPU size can hold, up to the model's limit
Badge a model card | Model authors | A VRAM badge snippet the page builds for the model you load

How to set it up:

• A web browser with internet access, so the page can read the model's configuration from the Hub
• No token is needed for public models
• Python and pip, only for the download step

1. Install the huggingface_hub package.
2. Download the Space into a local folder named llm-vram-calculator.
3. Move into that folder.
4. Open index.html in your web browser.

pip install -U huggingface_hub
hf download donaldJJ/llm-vram-calculator --repo-type space --local-dir llm-vram-calculator
cd llm-vram-calculator

How to use it: Open the page, enter a model id such as Qwen/Qwen3.8-27B and press Load. Then pick a quantization and a context length. The page can also open with a model already loaded when the page link ends in ?model= followed by a repository id.

Tool | What it does
Weights | Parameter count times bits per weight, divided by 8
KV cache | Counts only layers that keep a cache, in FP16, Q8_0 or Q4_0
Overhead | Adds 0.5 GiB plus 10 percent of weights and cache
GPU fit | Compares the total with GPU sizes and leaves 0.5 GiB free

Example: You enter Qwen/Qwen3.8-27B and press Load. At Q4_K_M with 8K context, the page shows 18.3 GiB, because only 16 of its 64 layers keep a KV cache. It says a 24 GB GPU holds up to 83K context.

For newcomers and experts: If you are new, VRAM is the memory on your graphics card, and a model must fit in it to run well. Enter a model id and read the GPU sizes that fit.

If you know the field, the KV cache counts full-attention layers, windowed layers with the window plus 512 tokens, and the compressed latent of MLA. Linear-attention and state-space layers add nothing. Every MoE expert is counted, with no CPU offload.

Enter a Hugging Face model id
The page reads the config in your browser
Weights, cache and overhead are added
GPU sizes that fit are listed

Title: Bits per weight used for the weights estimate
Q4_K_M: 4.84 bits
Q8_0: 8.5 bits
FP16: 16 bits

Good to know:

• Treat the results as rough guides, since serving software, CUDA graphs and image encoders can push real use higher
• Weights are sized from an assumed bit count per weight, so they differ from a real quantized file on disk
• Models behind an access request cannot be read by the page, so enter their numbers yourself
• License: mit
• Cost to run: none named
• Data it reaches: the Hugging Face Hub, to read a model's config and parameter count

On Hugging Face, 13 November 2026

• Likes: 3
• Built with: static
• License: mit

Links

• Space: https://huggingface.co/spaces/donaldJJ/llm-vram-calculator
• Files: https://huggingface.co/spaces/donaldJJ/llm-vram-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.