VRAM calculator

Pick a model, a quantisation and a context length: see how much GPU memory it needs and the cheapest GPUs that run it.

GPU memory needed17.2 GB

Cheapest GPUs that run it

GPUs from the catalog the model fits on, by price per hour.

The price is the cheapest GPU of the model × the GPU count. A 2–8 GPU server can cost a little more per GPU.

1 GB here is 1024³ bytes, like GPU memory. Hugging Face lists decimal gigabytes, so its numbers are 7% larger.

Weights are the real size of the published file: GGUF builds by unsloth and others, FP8 and BF16 from Hugging Face. Without a file we count from the parameters: Q4_K_M is 4.84 bits per weight, Q8_0 is 8.5.
The context cache follows the model's config.json: layers, KV heads and their size. Sliding windows (Gemma, GPT-OSS), compressed MLA caches (DeepSeek, Kimi) and linear-attention layers (Qwen3.5, Qwen3.8), whose cache does not grow with the context, are taken into account.
Runtime is the CUDA context and buffers: about 0.75 GB plus the logits buffer of a 512-token batch.
A model fits when the total is under 97.5% of the card: the driver and CUDA take the rest. vLLM takes 90% of the memory for its cache by default, which is expected.

Image and video models are counted for ComfyUI: the text encoder runs first and is swapped out, then the model, the VAE and the sampling memory stay.
“Minimum” is what ComfyUI does by default when memory is short. “No swapping” keeps everything loaded, so repeated runs are faster.
Working memory is our estimate for the listed output size and frame count. ComfyUI also runs the model on a smaller card by swapping part of the weights to system RAM, only slower.