Which AI models can your PC run?

How the estimate works

Memory needed is the model's weights (its parameter count from Hugging Face times the average bits per weight of the chosen llama.cpp quantization), plus the KV cache for the conversation, plus about 0.8 GB of runtime overhead. The KV cache is computed from each model's published config.json: grouped-query attention, DeepSeek-style latent attention, sliding-window layers (capped at the window) and linear-attention layers (no per-token cache) are all handled. 128 of the 145 models publish their config; for the rest we show the weights only.

Fast memory is a graphics card's video memory minus 0.6 GB, 75% of unified memory (the default GPU limit on Macs, often adjustable), or system RAM minus 4 GB for the operating system. If a model does not fit on a graphics card, the overflow can run from system RAM, which is slower.

Speed is limited by memory bandwidth: each generated token reads the active weights once. We scale the theoretical rate by the share real hardware achieves, measured on the AIPCs test bench, with a correction for AMD and Intel cards from published llama.cpp results. For mixture-of-experts models the active share comes from the expert layout in config.json; when that layout cannot be read reliably, we show whether the model fits but not a speed. Treat speeds as a range, not a promise: drivers, software and prompt length all move them.

These are estimates computed from sourced inputs, not measurements. See the methodology for how the underlying data is collected.