llmspeed is a free, open-source web tool: pick a GPU or Mac, a model and a quantization, and it types out a response at the estimated speed, tells you whether the model fits, and lets two setups race. How to use it, how to read its measured/derived/projected labels, and speed and time-to-first-token tables from the RTX 3060 to the 5090, Macs and the Ryzen AI Max+ 395.
