Choose a model, a quantization level and your hardware. The checker estimates the memory the model needs and tells you whether it fits entirely in GPU memory, needs to spill into system RAM, or will not run.
Can My PC Run It?
How the estimate works
Memory needed ≈ model weights (parameters × bits per weight ÷ 8) + KV cache for your context length + about 1 GB runtime overhead. For models whose layer layout we have recorded, the KV cache is computed from the architecture; for others we use a generic allowance and say so. Speed is shown only as a theoretical ceiling (memory bandwidth ÷ active weight size); it is not a benchmark. Full details: methodology.
FAQ
Is the result guaranteed?
No. Runtimes, drivers and settings change real memory use. Treat results within about 10% of your memory limit as borderline.
Why does a mixture-of-experts model need so much memory but run fast?
All experts must be loaded, so memory follows total parameters, but each token only uses the active parameters, so speed follows the active size.