Local AI PC sizer

Pick a model size, a precision, a context length and how many people use the model at once. The numbers under What it needs change as you pick, and the tower rebuilds itself: cards, memory sticks and power supplies slide in or out. Pick a graphics card from the list to build with it. Hover over or tap a part of the tower to see its line in the build list, or press a line to light up the part. Drag the tower to turn it and pinch to zoom; with a mouse, click the tower once and the scroll wheel zooms. With the tower focused, the arrow keys turn it, plus and minus zoom and 0 resets the view. Press Open the side panel to swing the glass away. Copy link copies a link that opens this same build.

The weights take the parameter count times the bits per parameter. 8-bit and 4-bit formats keep a scale for each small block of weights, so this page counts them as 8.5 and 4.5 bits. The KV cache holds a key and a value for every token of context in every layer, at 16-bit. Each size class has the layer count and attention heads of the common open models of its size. Each person gets a full context of their own, and each card keeps about 1 GB for its runtime. A model split across cards puts an equal share on each.Generating one token reads every active weight once and the whole KV cache once, so the speed is at most the memory bandwidth divided by the bytes read per token. The figure takes every context as full, as at the end of a long conversation; early in a conversation the cache is smaller and the ceiling higher. A model split across cards by layers runs on one card at a time, so the ceiling is that of one card. People served together share each step, which gives every one of them a token and reads a dense model's weights once for all of them. A mixture of experts keeps all its weights in memory, and its 235B class sends each token to 8 of 128 experts. People served together share its dense weights, about 7.8B, but each token picks its own experts, so 16 people read nearly two thirds of the experts in every step, not one sixteenth.System memory holds the model file while it loads, plus 8 GB for the system, and is never less than the graphics memory in the build. The power adds each card's board power, about 125 W for a desktop processor or 350 W for a workstation one, and the board, memory, drives and fans; the supply is sized a quarter above that. Three or four cards, or eight memory sticks, need a workstation board, which gets four or eight sticks so that every memory channel is filled. For the Apple chips the estimate leaves a quarter of the memory to macOS and apps.The specs are each maker's figures for memory, memory bandwidth and board power, read from the maker's own pages on 23 October 2026. Memory is in binary gigabytes, the unit memory is sold in; bandwidth is in decimal gigabytes per second, as the makers state it, and the formula counts bytes the same way.