AI Model Radar – measured hardware fit for open models (real GGUF sizes + per-architecture KV cache), with a JSON API and MCP server

Small update: the fit numbers now have a picture.

The card, seen through is a Space that draws a graphics card like an X-ray and runs the model you choose inside it: the Q4 GGUF file settles into memory, the KV cache grows as you drag the context from 2K to 128K, and when the card is full the rest spills into system RAM. 49 open models, 24 cards from 8 GB to 32 GB.

File sizes are measured from the published GGUF files, the context memory is computed from each model’s attention shape, the runtime share is an assumption. It shows no speed, because none is measured. If a number looks wrong for a model you run, I would like to hear it.