# LLM VRAM calculator: run it free, or from your agent

GPU memory an LLM needs, split into weights and KV cache, and which common GPUs hold it.

## Prompt

Use looot to check if I can run [model] on my [GPU] at [context] tokens. 1. Set up https://looot.ai/skill.md. 2. Read the model's parameter count, layers, KV heads and head size from its config. 3. Work out weights plus KV cache for FP16, INT8 and Q4. 4. Say which format fits, and compare with the price of running the same model through a chat completion endpoint in the catalog.

## What it needs

Parameters, weight format, context length, batch size and the model's layer shape

## What comes back

- total GPU memory in GB
- weights in GB
- KV cache in GB
- which GPUs fit and how many cards

## Endpoints and providers

None. The tool on the web page runs in the browser and calls no endpoint.

## What it costs

Nothing. The tool on the web page runs in the visitor's browser: no account, no API call and no limit per hour. The prompt above runs the same job from your own looot account with current data, billed per call.

## Buy the right GPU

See how much memory a model needs before you rent or buy a card.

## Pick a quantization

Compare FP16, INT8 and Q4 for the same model and context.

## Run models locally

Check if an open model such as a Gemma 4 fine-tune fits your machine.

## Size long context

See how the KV cache grows as you raise the context length or the batch.

## How do I calculate the VRAM an LLM needs?

Weights take parameters times bits per weight divided by 8. The KV cache takes 2 times layers times KV heads times head size times context times batch times bytes per value. Add an overhead share for the runtime. The calculator shows each part.

## How much VRAM does a 7B model need?

Pick Qwen2.5 7B Instruct and switch the weight format. The weights alone are about 15 GB at FP16 and about 4.3 GB at Q4_K, before the KV cache and overhead.

## Where do the model numbers come from?

The presets use the parameter count and the config.json of each model on Hugging Face, read 2026-10-14. Choose Custom to type any model's numbers.

## Why is my real usage lower?

The calculator counts every layer at full context. Models with sliding window layers, such as the Gemma family, keep a shorter cache in most layers, so they use less.

## Can I run the open Humanizer model?

Humanizer by jialinyyzz is a Gemma 4 12B fine-tune with about 11.9 billion parameters by its GGUF header. Choose Custom, type 11.9 and a weight format to see the weights. Read its layer numbers from the model card for the cache.

## What overhead should I use?

It depends on the runtime, so the field is yours to set. Start at the default and check against the memory your runtime reports.

## Is it free?

Yes. It runs in your browser, needs no account and has no limit.

## Related tools

- /free-tools/ai-model-comparison
- /free-tools/llm-token-cost-calculator
- /free-tools/ai-humanizer

## Last reviewed

2026-10-14.