Free · offline · no sign-up
LLM VRAM Calculator
Estimate the GPU memory needed to run a local language model — then see which 8, 12, 16, 24, 48 and 80 GB cards can carry it.
Everything is calculated in your browser.
Size your model
Results update as you type. No upload, no round-trip.
How the estimate works
Three memory pools decide whether a model launches — or spills into system RAM.
Parameters × bits
Every billion parameters costs roughly 4.8 GB at Q4_K_M. FP16 needs over three times as much room.
Context has a cost
Attention keys and values grow linearly with tokens. Long-context chat can consume more memory than the weights.
Leave breathing room
CUDA context, compute buffers and allocator fragmentation add about 0.75 GB plus 5% of weight memory.
A fit is not a guarantee
Windows, desktop compositing and other apps also need VRAM. Treat 90% utilization as the practical ceiling. A model that technically fits with 200 MB free will often stutter or fail at load.
of a GPU's memory should be your estimated total. The tier cards mark anything above that as tight.
Frequently asked questions
Why is my real usage higher than this estimate?
Your GPU driver, display, other apps and the inference engine's allocator all take memory. Batch size, tensor-parallel splits and multimodal projectors can also raise usage.
Does quantization affect quality?
Yes. Q8 is close to lossless, Q4_K_M is the usual quality/size tradeoff, and Q2/Q3 can visibly degrade reasoning and long outputs. The estimate uses average bits per weight.
What about multi-GPU setups?
Tensor parallelism approximates the memory of one card multiplied by the GPU count, but communication buffers and uneven layer splits add overhead. Keep extra headroom.
Is anything sent to a server?
No. The calculator is static HTML and vanilla JavaScript. Your inputs and results stay in the browser, and the page works offline once loaded.